Dokončenie GraphRAG evaluácie

README.md

@@ -15,7 +15,7 @@ FTS5/BM25 + embeddingy + RRF. Nad týmto baseline sa vo vetve GraphRAG
 dopĺňa knowledge graph uložený v Neo4j. Aktuálne je implementovaná
 grafová infraštruktúra, deterministický builder, vizualizácia grafu a
 samostatné grafové vyhľadávanie. Grafová vrstva má zdrojovo overený benchmark a zmrazenú opravenú
-verziu v6 s regresiou 320 historických otázok, kontrolou všetkých vrátených
+verziu v7 s novým TESTom 80 kombinácií a regresiou 320 historických otázok, kontrolou všetkých vrátených
 typovaných faktov a samostatným overením generovania odpovedí cez OpenWebUI API. Ďalším krokom je natívne zapojenie grafového
 nástroja a hlavného agenta do OpenWebUI.
 
@@ -870,17 +870,22 @@ grafové vyhľadávanie je implementované v `tools/graphrag/graph_search.py`.
 GraphRAG zatiaľ nezasahuje do stabilného klasického `/rag` toku; ďalšia
 etapa spojí grafové výsledky s klasickým RAG a OpenWebUI nástrojom.
 
-Aktuálna evaluácia GraphRAG **v6** má DEV **722/722** a historickú regresiu
-**320/320**, vrátane kontroly všetkých vrátených typovaných faktov oproti raw
-Markdownom. Generovanie sa meria samostatne: 42 DEV a finálna regresia 320
-historických otázok vrátane týchto 42; nejde o nový nezávislý TEST.
-Strict výsledok generovania je **316/320 (98,75 %)**. Celá sada vrátane
-živých testov má **565 úspešných testov**, bez preskočených prípadov.
-V6 prenáša úplnosť a verziu dát do kontextu, kontroluje všetky citované URL
-oproti podkladom konkrétneho dotazu a spája názov/rok s tagom tej istej osoby.
-Priame čítanie zdrojov opravilo 15 očakávaní; v1–v5.1 ostávajú zachované.
-Finálne metriky, prvotné odpovede aj opravné pokusy a obmedzenia sú v
-[finálnej správe v6](tools/graphrag/evaluation/results/final_v6_2026-10-25/REPORT.md).
+Aktuálna evaluácia GraphRAG **v7** má DEV **1 046/1 046** a nový TEST
+**80/80** pre grafové vyhľadávanie aj finálne odpovede. Modelový DEV má
+**46/46**; pri novom TESTe prešlo na prvý pokus **77/80 (96,25 %)**,
+po jednej povolenej oprave **80/80 (100 %)**. Prirodzený text sa skladá
+z overených štruktúrovaných tvrdení. Sú to nové kombinácie nad známym
+korpusom a šablónami, nie neznáme osoby ani ľudsky anotovaný benchmark.
+Samostatná historická regresia má **319/320 (99,69 %)**; nepodložená
+odpoveď v zostávajúcom prípade bola odmietnutá pred doručením.
+
+Všetkých **591 testov** vrátane živých prešlo. V7 opravuje ignorovanie rokov
+mimo 20xx, kontrolu rozporných tvrdení o úplnosti, počtov, povinných tvrdení
+a citácií vrátane referenčného Markdownu a HTML. Začiatok štúdia vyžaduje
+explicitný zdroj, neodvodzuje sa z adresára; opravilo sa 10 očakávaní.
+Všetky v1–v6 merania ostávajú zachované.
+[Finálna správa v7](tools/graphrag/evaluation/results/final_v7_2026-10-26/REPORT.md)
+uvádza presný protokol, opravy, regresiu aj obmedzenia.
 
 Historická v4 mala grafový strict TEST 320/320, doplnkovú kontrolu kontextových
 faktov 319/320 a LLM TEST 281/300 (93,67 %). Pôvodné otázky aj výsledky
@@ -1108,9 +1113,10 @@ spoľahlivého zdroja.
 
 ### 2. GraphRAG evaluácia
 
-V6 má zmrazený graf a úplnú historickú regresiu **320/320**.
-[Finálna správa](tools/graphrag/evaluation/results/final_v6_2026-10-25/REPORT.md)
-oddeľuje retrieval a generovanie odpovedí; pôvodné v1–v5.1 sa nemenia.
+V7 má zmrazený graf, DEV **1 046/1 046** a nový TEST **80/80**.
+[Finálna správa](tools/graphrag/evaluation/results/final_v7_2026-10-26/REPORT.md)
+oddeľuje nový TEST od historickej regresie a vyhľadávanie od modelu
+s kontrolami a rendererom; pôvodné v1–v6 sa nemenia.
 Pred formálnym porovnaním s klasickým RAG potrebujeme spoločnú metodiku
 a ideálne otázky anotované nezávislým človekom. Nový nezávislý TEST vyžaduje
 nové otázky; opakované meranie odhalených otázok je regresia.

docs/graphrag_v7_2026-10-26_worklog.md

@@ -0,0 +1,133 @@
+# Uzavretie bodu 3 — GraphRAG v7, 26. 10. 2026
+
+V7 opravuje konkrétne chyby z posledného auditu. Historické otázky,
+manifesty, tagy a výsledky v1–v6 zostávajú zachované. Záloha projektu,
+indexových súborov a grafu je v `/tmp/zp-v7-backup-20261026/`.
+
+## Opravy
+
+1. Dotaz rešpektuje ľubovoľný štvormiestny rok vrátane 1899, 1999, 2100
+   a 9999. Nepodporovaný zápis po „roku“, napríklad 10000 alebo 20xx,
+   sa odmieta cez `unsupported_year_constraint`. Podmienka sa nepremení
+   na vyhľadávanie bez časového obmedzenia.
+2. Upozornenie na neúplnosť už nepovoľuje kladné tvrdenie o úplnom zozname
+   v inej vete. Kontrola rozlišuje aj pravdivé popretie „Nie je to úplný
+   zoznam“. Testy obsahujú obe možnosti aj skloňované „všetkých študentov“.
+3. Model vyberá štruktúrované `claims`, `source_urls`, `status`, `count`;
+   `answer` ponecháva prázdne. Text pre používateľa zostaví
+   `render_graph_answer()` z overených hodnôt, bez gold a bez nových faktov.
+   Pôvodná HTTP odpoveď, parsované tvrdenia a vykreslená odpoveď ostávajú
+   osobitne uložené. Voľná próza modelu sa nepoužíva ako doručený text.
+4. Pri kontrole starého alebo neočakávaného voľného textu sa odmietajú
+   nepodložené číselné údaje. URL a doslovné hodnoty overených titulov sa
+   pred týmto porovnaním odlíšia od nových tvrdení. Prípad s dopísaným
+   rokom 2099 tak neprejde. Samotné porovnávanie čísel nie je úplný
+   sémantický analyzátor slovenčiny; doručený grafový text je preto
+   deterministický a obsahuje iba overené polia.
+5. `count` musí mať podklad v úplnej agregácii. Pozitívna faktová odpoveď
+   musí obsahovať tvrdenia a citácie; kontrolujú sa zdroje jednotlivých
+   uvedených atribútov. Prázdna úplná agregácia s počtom 0 je platná.
+6. Citácie sa kontrolujú v bežnom texte, inline a referenčnom Markdowne
+   aj HTML `href`/`src`, vrátane relatívnych URL, HTML entít a vnorených
+   zátvoriek v platnej URL. Povolené sú iba URL dodané pre konkrétny dotaz.
+7. Ročník adresára sa prestal používať ako začiatok štúdia. Builder aj
+   nezávislý oracle vyžadujú explicitnú vetu v texte. Graf uvádza
+   `start_year_basis=explicit_text|missing` a
+   `start_year_confidence=documented|unknown`. Oracle vedie `cohort_year`
+   oddelene; táto hodnota sa nevydáva za začiatok štúdia. Pri chýbajúcom
+   údaji nástroj vráti `missing_documented_start_year`.
+8. Hodnotenie mena povoľuje obmedzené gramatické prípony v prirodzenom
+   texte, ale štruktúrovanú identitu naďalej porovnáva presne. „Martina
+   NaN1“ sa nezamieňa za iného Martina. Pri deterministickom texte sú
+   mená aj tak v pôvodnom tvare. Parser JSON neprepisuje neplatné hodnoty;
+   jeden spoločný opravný požiadavok pre formát/kontext ostáva limitom.
+9. Evaluačné CLI vyžadujú explicitný dataset, audit a manifest. Odpovedný
+   runner vyžaduje aj zoznam DEV; historická regresia navyše vlastné ID.
+   Už nevyberajú v4 ako tichý predvolený benchmark. Hashované cesty kódu
+   generovania sú relatívne voči projektu, aby sa dali overiť aj inde.
+
+## Benchmark a oddelenie meraní
+
+Všetkých 1 042 odhalených v6 otázok prešlo do DEV. Pridali sa štyri
+potrebné regresné otázky na roky mimo 20xx. DEV má **1 046 otázok**.
+Odstránenie odhadu štúdia opravilo 10 očakávaní; údaje, ktoré dokument
+neuvádza, sa očakávajú ako nenájdené. ID ani text pôvodných otázok sa
+nevymazali. Zmeny sú v `v7_preparation/source_corrections.json`.
+
+Nový **TEST má 80 otázok**, 72 kladných a 8 záporných, vo všetkých
+desiatich kategóriách. Seed, kvóty a výber boli uložené pred vyhľadávaním.
+Žiadna logická query skupina sa neprekrýva s odhalenými otázkami ani DEV.
+Ide o nové kombinácie nad známou wiki, osobami a šablónami. Nie je to
+experiment s neznámymi osobami, novým korpusom alebo ľudskou anotáciou.
+
+Počet otázok v kategóriách je obmedzený dostupnými novými skupinami:
+spojenie autor–kategória má dve nové skupiny, výpis kategórií sedem.
+Kvóty sa neurčovali podľa úspechu retrievera. Reprezentatívnu vzorku
+20 otázok (dve z každej kategórie) skontrolovala AI priamo oproti
+Markdownom pred meraním. `source_review_packet.json` obsahuje pôvodné
+výňatky a dôkazy; `source_review.csv` závery. **Človek ju neanotoval.**
+Odstránenie odhadu z adresára teda rieši neistotu z auditu; nenahrádza
+nezávislú ľudskú anotáciu celého korpusu.
+
+Graf sa zmrazí po kompletnom DEV. Generovanie má vopred deklarovaných
+46 DEV otázok: predchádzajúcich 42 a štyri nové rokové regresie.
+Finálny nový TEST sa spustí raz po úspešnom DEV. Výsledky sa nepoužijú
+na ladenie tejto verzie. Všetkých 320 pôvodných historických otázok sa
+meria oddelene ako regresia; jej prekryv s modelovým DEV je priznaný.
+
+## Zmrazenie a nasadenie
+
+`frozen_v7.json` viaže implementáciu, raw Markdown, dataset, JSON vstupy,
+graf, publikačnú politiku a parametre. Deklarácia `frozen_answers.json`
+navyše fixuje model, prompt, renderer a presné ID merania. Kompletný
+`frozen_release_v7.json` pokrýva všetky runtime moduly a merané artefakty.
+
+Docker sa overuje cez `/health`, reálne `/rag`, grafové pozitívne aj
+negatívne dotazy, novú rokovú regresiu a chýbajúci začiatok štúdia.
+JSON, SQLite a Neo4j majú spoločnú generáciu po `/sync` s `pull_git=false`.
+Publikačná politika ostáva: iba explicitné `published=false` sa vylúči.
+
+## Príkazy
+
+Konfiguráciu načítajte z prostredia bez jej vypisovania. Pri reprodukcii
+vždy používajte nové výstupné cesty; existujúce otázky a merania neprepisujte.
+
+```bash
+# Jednorazové vyhlásenie benchmarku; hotový dataset neregenerovať.
+python -m tools.graphrag.evaluation.benchmark_v7
+
+RUN_NEO4J_TESTS=1 RUN_LIVE_TESTS=1 RUN_LIVE_RAG_E2E=1 \
+  HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 python -m pytest -q
+
+python -m tools.graphrag.evaluation.evaluate_graphrag \
+  --questions tools/graphrag/evaluation/graphrag_questions_v7.json --split dev \
+  --audit tools/graphrag/evaluation/results/v7_preparation/source_audit.json \
+  --output-dir NOVY_GRAF_DEV
+python -m tools.graphrag.evaluation.freeze_graphrag \
+  --questions tools/graphrag/evaluation/graphrag_questions_v7.json \
+  --dev-results NOVY_GRAF_DEV/graphrag_dev.json \
+  --audit tools/graphrag/evaluation/results/v7_preparation/source_audit.json \
+  --purpose independent_holdout --output NOVY_MANIFEST.json
+python -m tools.graphrag.evaluation.evaluate_graphrag \
+  --questions tools/graphrag/evaluation/graphrag_questions_v7.json --split test \
+  --audit tools/graphrag/evaluation/results/v7_preparation/source_audit.json \
+  --freeze-manifest NOVY_MANIFEST.json --output-dir NOVY_GRAF_DEV
+
+python -m tools.answer_evaluation.evaluate_graph_answers \
+  --questions tools/graphrag/evaluation/graphrag_questions_v7.json \
+  --audit tools/graphrag/evaluation/results/v7_preparation/source_audit.json \
+  --freeze-manifest NOVY_MANIFEST.json \
+  --dev-ids tools/graphrag/evaluation/results/v7_preparation/generation_dev_ids.json \
+  --split dev --output-dir NOVY_LLM_DEV
+```
+
+Pre nový TEST použite tie isté parametre, `--split test`,
+`--dev-results NOVY_LLM_DEV/answers_dev.json` a nový výstupný adresár.
+Pri historickej regresii použite `--split regression` a navyše
+`--regression-ids tools/graphrag/evaluation/results/v7_preparation/regression_ids.json`.
+Opakovaný beh už odhaleného TESTu je reprodukcia/regresia, nie ďalší
+nezávislý experiment. Po ladení podľa TESTu treba nový vopred deklarovaný TEST.
+
+Kompletné výsledky a zostávajúce metodické obmedzenia sú v
+[správe v7](../tools/graphrag/evaluation/results/final_v7_2026-10-26/REPORT.md).
+Ďalší blok je natívna integrácia nástrojov a hlavného agenta v OpenWebUI.

docs/vystupny_kontrakt.md

@@ -1,5 +1,9 @@
 # Spoločný výstupný kontrakt RAG a GraphRAG
 
+V7 rozširuje kontrolu formátov citácií a rozporov v úplnosti. Grafové
+generovanie má validované štruktúrované tvrdenia a deterministický renderer.
+[Aktuálny protokol](graphrag_v7_2026-10-26_worklog.md).
+
 V6 prenáša polia úplnosti a verzie aj do kontextu LLM a kontroluje ich pri
 generovaní. Kontrola citácií sa vzťahuje na zoznam aj odkazy v texte odpovede.
 [Implementácia, meranie a obmedzenia](graphrag_v6_2026-10-25_worklog.md).

test/test_bod3_closure.py

@@ -0,0 +1,113 @@
+import copy
+import json
+
+import pytest
+
+from tools.answer_evaluation.context_policy import answer_context_issues, citation_check, coverage_issues, prose_issues, text_contains_value
+from tools.answer_evaluation.checked_generation import generate_checked_answer
+from tools.answer_evaluation.render_answer import render_graph_answer
+from tools.graphrag.graph_search import extract_query_constraints
+from tools.graphrag.graphrag import extract_start_year
+
+
+@pytest.fixture
+def context():
+    return {'status': 'ok', 'complete': True, 'truncated': False, 'aggregation': None,
+            'sources': [{'source_url': 'https://example.test/students/2018/eva'}],
+            'facts': [{'person': 'Eva Nová', 'year': 2026, 'title': 'Práca o 3 modeloch',
+                       'source_url': 'https://example.test/students/2018/eva'}], 'answer_candidates': []}
+
+
+@pytest.fixture
+def answer(context):
+    return {'answer': '', 'status': 'ok', 'claims': [{'person': 'Eva Nová', 'year': 2026, 'title': 'Práca o 3 modeloch'}],
+            'source_urls': [context['sources'][0]['source_url']], 'count': None}
+
+
+@pytest.mark.parametrize('year', [1899, 1999, 2099, 2100, 9999])
+def test_year_constraint_is_never_silently_dropped(year):
+    assert extract_query_constraints(f'Aká práca je v roku {year}?')['years'] == [year]
+
+
+@pytest.mark.parametrize('text', ['Toto je úplný zoznam. Výpis je neúplný.',
+                                'Uvádzam všetkých študentov. Výpis je neúplný.',
+                                'Úplný zoznam, ale dôkazy sú orezané.'])
+def test_caveat_does_not_license_contradictory_exhaustive_claim(text):
+    assert 'unsupported_exhaustive_statement' in coverage_issues({'answer': text, 'status': 'ok'}, {'complete': False})
+
+
+@pytest.mark.parametrize('text', ['Výpis je neúplný.', 'Toto nie je úplný zoznam.', 'Nie je to úplný zoznam.'])
+def test_truthful_scope_notice_is_not_rejected(text):
+    assert coverage_issues({'answer': text, 'status': 'ok'}, {'complete': False}) == []
+
+
+@pytest.mark.parametrize('text', ['[zdroj][s]\n\n[s]: /vymysleny', '<a href="/vymysleny">zdroj</a>',
+                                '<img src="//evil.test/x">', '[s]: javascript:alert(1)',
+                                '<a href="https:&#47;&#47;evil.test/x">x</a>'])
+def test_all_citation_syntaxes_are_checked(text):
+    assert not citation_check({'answer': text, 'source_urls': []}, {'https://example.test/a'})['citations_ok']
+
+
+def test_legitimate_balanced_parentheses_and_reference_html_links():
+    url = 'https://example.test/a_(b)'
+    text = f'[x]({url}) a {url}.\n[s]: <{url}>\n<a href="{url}">x</a>'
+    assert citation_check({'answer': text, 'source_urls': [url]}, {url})['citations_ok']
+
+
+def test_count_requires_explicit_complete_aggregation(context, answer):
+    answer['count'] = 98765
+    assert 'unsupported_total_count' in answer_context_issues(answer, context, check_text=False)
+
+
+@pytest.mark.parametrize('field,issue', [('claims', 'missing_required_claims'), ('source_urls', 'missing_required_citations')])
+def test_positive_answer_needs_facts_and_citations(context, answer, field, issue):
+    answer[field] = []
+    assert issue in answer_context_issues(answer, context, check_text=False)
+
+
+def test_every_claim_field_needs_its_source(context, answer):
+    context['facts'].append({'person': 'Eva Nová', 'keyword': 'python', 'source_url': 'https://example.test/child'})
+    context['sources'].append({'source_url': 'https://example.test/child'})
+    answer['claims'][0]['keyword'] = 'python'
+    assert 'missing_claim_source' in answer_context_issues(answer, context, check_text=False)
+    answer['source_urls'].append('https://example.test/child')
+    assert answer_context_issues(answer, context, check_text=False) == []
+
+
+def test_wrong_extra_prose_year_is_rejected_but_citation_path_year_is_not(context, answer):
+    answer['answer'] = 'Eva Nová má prácu z roku 2026. ' + answer['source_urls'][0]
+    assert prose_issues(answer) == []
+    answer['answer'] += ' Obhajoba bola v roku 2099.'
+    assert prose_issues(answer) == ['unsupported_prose_number']
+
+
+def test_renderer_uses_only_validated_fields_and_preserves_model_reply(context, answer):
+    original = copy.deepcopy(answer)
+    def complete(_): return {'choices': [{'message': {'content': json.dumps(answer)}}]}
+    result = generate_checked_answer('Názov a rok?', context, 'Pokyny', complete)
+    parsed_original = {key: value for key, value in original.items() if value is not None}
+    assert result['accepted_for_delivery'] and result['model_answer'] == parsed_original
+    assert result['attempts'][0]['answer'] == parsed_original
+    assert result['attempts'][0]['response']['choices'][0]['message']['content'] == json.dumps(original)
+    assert '2026' in result['answer']['answer'] and 'Práca o 3 modeloch' in result['answer']['answer']
+    assert answer == original
+
+
+def test_zero_count_is_valid_without_claims_or_sources():
+    context = {'status': 'ok', 'facts': [], 'sources': [], 'complete': True,
+               'aggregation': {'count': 0, 'complete': True}}
+    answer = {'answer': '', 'status': 'ok', 'claims': [], 'source_urls': [], 'count': 0}
+    rendered = render_graph_answer(answer, context)
+    assert answer_context_issues(rendered, context) == []
+
+
+def test_directory_cohort_does_not_become_study_start():
+    assert extract_start_year('pages/students/2016/eva/README.md', [{'text': 'Diplomová práca 2026'}]) is None
+    assert extract_start_year('pages/students/2016/eva/README.md', [{'text': 'Rok začiatku štúdia: 2017'}]) == 2017
+    assert extract_start_year('pages/students/2016/eva/README.md', [{'text': 'Beginning of the study: 2018'}]) == 2018
+
+
+def test_person_inflection_keeps_distinct_names_distinct():
+    assert text_contains_value('Adriána Remiáša a Martina NaN1', 'Martin NaN1', person=True)
+    assert not text_contains_value('Martin Jančura', 'Martin NaN1', person=True)
+    assert not text_contains_value('Martin Malinovský', 'Martin Malý', person=True)

tools/answer_evaluation/README.md

@@ -1,6 +1,29 @@
 # Generovanie odpovedí nad zmrazeným grafom
 
-Aktuálne meranie **v6** je historická regresia 320 známych otázok vrátane
+Aktuálne meranie **v7** používa 46 známych DEV otázok, nový query-group TEST
+80 otázok a oddelenú historickú regresiu 320 otázok.
+Model vracia štruktúrované tvrdenia, `answer` je prázdne. Aplikácia overí
+fakty, väzby, počty a citácie a vykreslí prirodzený text deterministicky.
+Voľná próza modelu sa ako finálna odpoveď nedoručuje. Pôvodné HTTP odpovede,
+parsované aj vykreslené odpovede a každý opravný pokus sú uložené.
+
+Nový TEST má prvý pokus **77/80 (96,25 %)** a finálny výsledok **80/80**;
+jednu opravu potrebovali tri otázky. Historická regresia má 319/320,
+s jedným odmietnutým nepodloženým akademickým názvom. Nejde o meranie voľnej jazykovej
+formulácie ani neznámeho korpusu. Modelový DEV má 46/46.
+[Správa v7](../graphrag/evaluation/results/final_v7_2026-10-26/REPORT.md).
+
+Všetky evaluačné vstupy treba zvoliť explicitne. Nový TEST používa
+`--split test`; známa regresia `--split regression --regression-ids ...`.
+Oba potrebujú dokončený `--dev-results` s rovnakými hashmi a modelom.
+Príkazy sú v [pracovnom zázname v7](../../docs/graphrag_v7_2026-10-26_worklog.md).
+Budúci OpenWebUI agent musí dodržať `accepted_for_delivery` a odovzdať
+overený text; nekontrolované prepísanie rendererovej odpovede by tieto
+záruky obišlo. Natívny výber nástrojov ostáva ďalším blokom.
+
+## Historická v6
+
+Meranie **v6** je historická regresia 320 známych otázok vrátane
 42 DEV otázok. [Výsledná správa](../graphrag/evaluation/results/final_v6_2026-10-25/REPORT.md)
 uvádza samostatne prvý pokus, výsledok po jednom povolenom opravnom požiadavku,
 odmietnuté odpovede a JSON chyby. Žiadna chyba sa nevynecháva z menovateľa.

tools/answer_evaluation/check_context_scenarios.py

@@ -67,7 +67,7 @@ def main():
     args = parser.parse_args()
     if args.output.exists():
         parser.error('Refusing to overwrite results.')
-    files = [Path(__file__), HERE/'graph_answer_prompt.txt', HERE/'context_policy.py', HERE/'checked_generation.py', HERE/'answer_format.py']
+    files = [Path(__file__), HERE/'graph_answer_prompt.txt', HERE/'context_policy.py', HERE/'checked_generation.py', HERE/'answer_format.py', HERE/'render_answer.py']
     hashes = {str(p): sha256_file(p) for p in files}
     declared = scenarios()
     args.output.parent.mkdir(parents=True, exist_ok=True)

tools/answer_evaluation/check_deployment.py

@@ -19,24 +19,28 @@ from tools.rag.common import DB_FILE
 def main():
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument('--output', type=Path, required=True)
+    parser.add_argument('--questions', type=Path, required=True)
+    parser.add_argument('--audit', type=Path, required=True)
+    parser.add_argument('--freeze-manifest', type=Path, required=True)
+    parser.add_argument('--release-manifest', type=Path)
     args = parser.parse_args()
     if args.output.exists(): parser.error('Refusing to overwrite deployment verification.')
     base = Path('tools/graphrag/evaluation')
-    release_path = base / 'frozen_release_v6.json'
+    release_path = args.release_manifest
     release_match = None
-    if release_path.exists():
+    if release_path is not None:
         release = json.loads(release_path.read_text())
         for path, digest in release['runtime_files'].items():
             assert sha256_file(Path(path)) == digest, f'Release code mismatch: {path}'
-        assert sha256_file(base/'frozen_v6.json') == release['graph_freeze_sha256']
+        assert sha256_file(args.freeze_manifest) == release['graph_freeze_sha256']
         release_match = True
-    freeze = json.loads((base/'frozen_v6.json').read_text())
+    freeze = json.loads(args.freeze_manifest.read_text())
     health = request_json('http://localhost:8000/health')
     assert health['ready'] and all(v['ready'] for v in health['tools'].values())
     with create_neo4j_driver() as driver, driver.session() as session:
         graph = graph_fingerprint(session)
-        verify_freeze(freeze, run_metadata(base/'graphrag_questions_v6.json'), graph,
-                      audit=base/'results/v6_preparation/source_audit.json', **freeze['parameters'])
+        verify_freeze(freeze, run_metadata(args.questions), graph,
+                      audit=args.audit, **freeze['parameters'])
         current = read_version(Path('data'))
         assert current == sqlite_version(DB_FILE) == graph_version(session) == health['index_version']
         documents = session.run('MATCH (d:Document {graph_scope:"zpwiki"}) RETURN d.path AS path, d.published AS published').data()
@@ -50,7 +54,10 @@ def main():
     for id, question, expected in [
         ('known_title', 'Aký je názov práce osoby Ján Holp a rok práce?', 'ok'),
         ('generic_label_not_title', 'Uveď názov práce osoby Dárius Lindvai v roku 2019 s kľúčovým slovom python.', 'no_answer'),
-        ('unknown_person', 'Akú prácu má osoba Jozef Neznámy Overovací?', 'no_answer')]:
+        ('unknown_person', 'Akú prácu má osoba Jozef Neznámy Overovací?', 'no_answer'),
+        ('outside_century', 'Aký je názov práce osoby Ján Holp v roku 1999?', 'no_answer'),
+        ('unsupported_year', 'Aký je názov práce osoby Ján Holp v roku 10000?', 'no_answer'),
+        ('unknown_start', 'Aký je rok začiatku štúdia osoby Tomáš Kuchárik?', 'no_answer')]:
         result = search_graph(question)
         assert result['status'] == expected and result['index_version'] == current
         cases[id] = {'status': result['status'], 'reason': result['reason'], 'complete': result['complete']}

tools/answer_evaluation/checked_generation.py

@@ -2,6 +2,7 @@
 import json
 from .answer_format import parse_answer, Answer
 from .context_policy import answer_context_issues
+from .render_answer import render_graph_answer
 
 
 def generate_checked_answer(question: str, context: dict, prompt: str, complete) -> dict:
@@ -29,10 +30,19 @@ def generate_checked_answer(question: str, context: dict, prompt: str, complete)
                 return result
             feedback='Vráť jediný platný JSON objekt podľa schémy, bez zmeny faktov z pôvodného kontextu. Schéma: '+json.dumps(Answer.model_json_schema(),ensure_ascii=False)
         else:
-            issues=answer_context_issues(answer,context)
+            issues=answer_context_issues(answer,context,check_text=False)
             record['context_issues']=issues
             result['answer']=answer
             if not issues:
+                rendered = render_graph_answer(answer, context)
+                delivery_issues = answer_context_issues(rendered, context)
+                if delivery_issues:
+                    issues = delivery_issues
+                    record['context_issues'] = issues
+                else:
+                    record['rendered_answer'] = rendered
+                    result['model_answer'] = answer
+                    result['answer'] = rendered
                     result['accepted_for_delivery']=True
                     return result
             if number==1:
@@ -41,6 +51,6 @@ def generate_checked_answer(question: str, context: dict, prompt: str, complete)
             feedback=('Odpoveď porušila pravidlá dodaného kontextu: '+', '.join(issues)+'. '
                 'Oprav jediný JSON objekt výhradne z pôvodných podkladov. Použi iba source_urls zo sources. '
                 'Typ práce nie je keyword. answer_candidates dokladajú spojenie požadovaného názvu alebo roku a tagu cez tú istú osobu. '
-                'Pri neúplnom výstupe uveď: Výpis je neúplný.')
+                'Answer ponechaj prázdny; text aj upozornenie zostaví aplikácia.')
         messages += [{'role':'assistant','content':text},{'role':'user','content':feedback}]
     return result

tools/answer_evaluation/context_policy.py

@@ -2,6 +2,8 @@
 from __future__ import annotations
 
 import re
+from html import unescape
+from html.parser import HTMLParser
 import unicodedata
 
 
@@ -10,12 +12,40 @@ def normalized(value) -> str:
     return ' '.join(re.findall(r'[a-z0-9]+',''.join(c for c in text if not unicodedata.combining(c)).casefold()))
 
 
+class _HTMLLinks(HTMLParser):
+    def __init__(self):
+        super().__init__(convert_charrefs=True)
+        self.links = set()
+
+    def handle_starttag(self, tag, attrs):
+        self.links.update(value for key, value in attrs if key in {'href', 'src'} and value)
+
+
 def text_urls(text: str) -> set[str]:
-    urls = re.findall(r'(?:[a-zA-Z][a-zA-Z0-9+.-]*://|www\.)[^\s<>"\x27`“”„«»‘’\]\)]+', text)
-    # Relative Markdown targets are citations too, and must be supplied by
-    # the tool just like absolute links. Never resolve/guess another URL.
-    urls += re.findall(r'\[[^\]]*\]\(\s*<?([^\s)>]+)>?', text)
-    return {url.rstrip('.,;:!?') for url in urls}
+    """Inspect absolute, inline, reference-style Markdown and HTML links."""
+    text = unescape(text)
+    urls = set()
+    for url in re.findall(r'(?:[a-zA-Z][a-zA-Z0-9+.-]*://|www\.)[^\s<>"\x27`“”„«»‘’\]]+', text):
+        url = url.rstrip('.,;:!?')
+        while url.endswith(')') and url.count(')') > url.count('('):
+            url = url[:-1]
+        urls.add(url)
+    for match in re.finditer(r'\]\(\s*', text):
+        tail = text[match.end():]
+        if tail.startswith('<'):
+            urls.add(tail[1:].split('>', 1)[0]); continue
+        level = 0; end = 0
+        for end, char in enumerate(tail):
+            if char.isspace() or (char == ')' and level == 0): break
+            if char == '(': level += 1
+            if char == ')': level -= 1
+        else:
+            end = len(tail)
+        if tail[:end]: urls.add(tail[:end])
+    for match in re.finditer(r'^\s{0,3}\[[^]\n]+\]:\s*(?:<([^>]+)>|(\S+))', text, flags=re.M):
+        urls.add(match[1] or match[2])
+    parser = _HTMLLinks(); parser.feed(text)
+    return urls | parser.links
 
 
 def citation_check(answer: dict, allowed: set[str]) -> dict:
@@ -56,30 +86,72 @@ def joined_candidates(facts: list[dict], intent: dict) -> list[dict]:
 def coverage_issues(answer: dict, context: dict) -> list[str]:
     incomplete = context.get('truncated') is True or context.get('complete') is False
     unknown = context.get('complete') is None
-    text=normalized(answer.get('answer',''))
-    caveat=any(phrase in text for phrase in ['vysledok je neuplny','vysledky su neuplne','vypis je neuplny',
-        'zoznam je neuplny','dokazy su orezane','vysledok je orezany','vysledky su orezane',
-        'nemusi byt uplny','nemusi byt uplna','nie je uplny','nie su uplne','nemozno zarucit uplnost'])
+    clauses = [normalized(part) for part in re.split(r'[.!?;\n]+', answer.get('answer', ''))]
+    caveat = any(re.search(r'(?:vypis|zoznam|vysled\w*|dokazy) (?:je |su |nie je |nie su )?(?:neupln\w*|orezan\w*)'
+                          r'|nie (?:je|su) (?:to )?upln\w*|nemozno zarucit uplnost|nemusi byt upln', clause)
+                 for clause in clauses)
     issues = []
     if incomplete and answer.get('status') == 'ok' and not caveat:
         issues.append('missing_incompleteness_notice')
-    exhaustive=any(phrase in text for phrase in ['uplny zoznam','kompletny zoznam','vsetci studenti','vsetky osoby','vsetky prace'])
-    if (incomplete or unknown) and exhaustive and not caveat:
+    if incomplete or unknown:
+        for clause in clauses:
+            for match in re.finditer(r'(?:upln\w*|kompletn\w*) (?:zoznam|vypis|prehlad)'
+                                     r'|vset(?:ci|ky|kych) (?:student\w*|osob\w*|prac\w*)', clause):
+                prefix = clause[:match.start()]
+                # Negation must qualify this particular assertion. A caveat in
+                # another sentence never licenses a positive exhaustive claim.
+                if not re.search(r'(?:nie (?:je|su)(?: to)?|nejde o|neuvadzam|'
+                                 r'nemozno (?:zarucit|uviest)|nemozem (?:zarucit|poskytnut))\s*$', prefix):
                     issues.append('unsupported_exhaustive_statement')
-    return issues
+    return list(dict.fromkeys(issues))
 
 
-def answer_context_issues(answer: dict, context: dict) -> list[str]:
+def prose_issues(answer: dict) -> list[str]:
+    """Conservative numeric consistency check; final prose is rendered from claims."""
+    text = str(answer.get('answer', ''))
+    for url in sorted(text_urls(text), key=len, reverse=True):
+        text = text.replace(url, '')
+    allowed_numbers = set()
+    for claim in answer.get('claims', []):
+        for key, value in claim.items():
+            if value is None: continue
+            if key in {'year', 'start_year'}: allowed_numbers.add(str(value))
+            else: text = re.sub(re.escape(str(value)), '', text, flags=re.I)
+    if answer.get('count') is not None: allowed_numbers.add(str(answer['count']))
+    unsupported = set(re.findall(r'\b\d+\b', text)) - allowed_numbers
+    return ['unsupported_prose_number'] if unsupported else []
+
+
+def text_contains_value(text: str, value: str, *, person: bool = False) -> bool:
+    value = normalized(value); text = normalized(text)
+    if not person: return value in text
+    suffix = r'(?:a|u|ovi|om|e|ou|y|ovej|oveho|ym|emu)?'
+    pattern = r'\b' + r'\s+'.join(re.escape(token) + (suffix if token.isalpha() else '')
+                                       for token in value.split()) + r'\b'
+    return bool(re.search(pattern, text))
+
+
+def answer_context_issues(answer: dict, context: dict, *, check_text: bool = True) -> list[str]:
     issues=[]
     if not citation_check(answer,allowed_urls(context))['citations_ok']:
         issues.append('unsupported_citation')
     facts=context.get('facts',[])
     aggregation=context.get('aggregation') or {}
-    if answer.get('count') is not None and aggregation:
+    if answer.get('count') is not None:
         if aggregation.get('complete') is not True:
             issues.append('unsupported_total_count')
         elif answer['count']!=aggregation.get('count'):
             issues.append('incorrect_context_count')
+    claims = answer.get('claims', [])
+    cited = set(answer.get('source_urls', []))
+    if answer.get('status') == 'ok' and facts and not claims:
+        issues.append('missing_required_claims')
+    if claims and not cited:
+        issues.append('missing_required_citations')
+    if answer.get('status') == 'no_answer' and claims:
+        issues.append('claims_on_no_answer')
+    if check_text and not str(answer.get('answer', '')).strip():
+        issues.append('empty_answer_text')
     for claim in answer.get('claims',[]):
         person=normalized(claim.get('person'))
         rows=[f for f in facts if normalized(f.get('person'))==person]
@@ -88,6 +160,12 @@ def answer_context_issues(answer: dict, context: dict) -> list[str]:
         for key in ['keyword','category','document_author','start_year']:
             if claim.get(key) is not None and not any(normalized(row.get(key))==normalized(claim[key]) for row in rows):
                 issues.append('unsupported_'+key)
+        for key in ['title', 'year', 'start_year', 'keyword', 'category', 'document_author']:
+            if claim.get(key) is None: continue
+            evidence = {row.get('source_url') for row in rows
+                        if normalized(row.get(key)) == normalized(claim[key]) and row.get('source_url')}
+            if evidence and not evidence & cited:
+                issues.append('missing_claim_source')
         academic={k:claim[k] for k in ['title','year'] if claim.get(k) is not None}
         if academic and not any(all(normalized(row.get(k))==normalized(v) for k,v in academic.items()) for row in rows):
             issues.append('unsupported_academic_binding')
@@ -105,5 +183,7 @@ def answer_context_issues(answer: dict, context: dict) -> list[str]:
             issues.append('split_or_missing_person_join')
     if context.get('status') in {'no_answer','ambiguous','error'} and answer.get('status')=='ok':
         issues.append('unsupported_positive_status')
+    issues.extend(prose_issues(answer))
+    if check_text or str(answer.get('answer', '')).strip():
         issues.extend(coverage_issues(answer,context))
     return list(dict.fromkeys(issues))

tools/answer_evaluation/evaluate_graph_answers.py

@@ -24,7 +24,7 @@ from tools.graphrag.evaluation.reproducibility import graph_fingerprint, run_met
 from tools.graphrag.evaluation.source_oracle import load_source_oracle, norm
 from tools.rag.evaluation.rag_runner import DEFAULT_MODEL, OPENWEBUI_URL, request_json
 from .checked_generation import generate_checked_answer
-from .context_policy import joined_candidates, citation_check, allowed_urls, answer_context_issues
+from .context_policy import joined_candidates, citation_check, allowed_urls, answer_context_issues, prose_issues, text_contains_value
 
 HERE = Path(__file__).resolve().parent
 
@@ -86,7 +86,10 @@ def grade_answer(case: dict, answer: dict, oracle: dict, *, context: dict | None
     policy_issues=answer_context_issues(normalized,context) if context is not None else []
     grade['context_issues']=policy_issues
     text = norm(answer.get('answer', ''))
-    missing = [v for v in case.get('expected_answer_contains', []) if norm(v) not in text]
+    people = {str(b['person']) for b in case['graph']['expected_bindings'] if b.get('person')}
+    missing = [v for v in case.get('expected_answer_contains', [])
+               if not text_contains_value(answer.get('answer', ''), str(v), person=str(v) in people)]
+    policy_issues.extend(prose_issues(normalized))
     # The initial sample exposed a real contradiction in otherwise correct
     # category claims. Check it explicitly rather than silently accepting prose.
     wrong_thesis_label = 'dizertac' in text and any(str(b.get('category', '')).startswith('dp')
@@ -101,14 +104,14 @@ def grade_answer(case: dict, answer: dict, oracle: dict, *, context: dict | None
 def main():
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument('--output-dir', type=Path, required=True)
-    parser.add_argument('--sample', type=Path, default=BASE / 'results/final_v4_2026-10-25/openwebui_sample.json')
     parser.add_argument('--workers', type=int, default=3)
     parser.add_argument('--model', default=DEFAULT_MODEL)
     parser.add_argument('--split', choices=['dev', 'test', 'regression'], default='test')
-    parser.add_argument('--questions', type=Path, default=BASE/'graphrag_questions_v4.json')
-    parser.add_argument('--audit', type=Path, default=BASE/'results/v4_preparation/source_audit.json')
-    parser.add_argument('--freeze-manifest', type=Path, default=BASE/'frozen_v4.json')
-    parser.add_argument('--dev-ids', type=Path)
+    parser.add_argument('--questions', type=Path, required=True)
+    parser.add_argument('--audit', type=Path, required=True)
+    parser.add_argument('--freeze-manifest', type=Path, required=True)
+    parser.add_argument('--dev-ids', type=Path, required=True)
+    parser.add_argument('--regression-ids', type=Path)
     parser.add_argument('--dev-results', type=Path, help='Required completed DEV before final regression.')
     args = parser.parse_args()
     if args.output_dir.exists():
@@ -117,19 +120,26 @@ def main():
         parser.error('OPENWEBUI_API_KEY missing.')
     if args.workers < 1:
         parser.error('--workers must be positive.')
-    if args.split=='regression' and args.dev_results is None:
+    if args.split in {'regression', 'test'} and args.dev_results is None:
         parser.error('Final regression requires completed --dev-results.')
     dataset=args.questions; audit=args.audit; frozen=args.freeze_manifest
     manifest = json.loads(frozen.read_text())
     questions = json.loads(dataset.read_text()); oracle = load_source_oracle()
-    sample = json.loads(args.sample.read_text()); exposed = set(sample['selected_ids'])
-    test = [q for q in questions if q['split'] == 'test' and q['id'] not in exposed]
-    assert len(exposed) == 20 and len(test) == 300
-    development=set(json.loads(args.dev_ids.read_text())) if args.dev_ids else exposed
-    selected = ([q for q in questions if q['split']=='test'] if args.split=='regression' else
-                test if args.split=='test' else [q for q in questions if q['id'] in development])
-    if args.split=='regression' and manifest.get('evaluation_kind')!='historical_regression':
-        parser.error('Regression requires an explicitly declared historical regression freeze.')
+    development=set(json.loads(args.dev_ids.read_text()))
+    test = [q for q in questions if q['split'] == 'test']
+    if args.split == 'regression':
+        if args.regression_ids is None: parser.error('Regression needs explicit --regression-ids.')
+        regression_ids = set(json.loads(args.regression_ids.read_text()))
+        selected = [q for q in questions if q['id'] in regression_ids]
+        if len(selected) != len(regression_ids): parser.error('Unknown/duplicate regression IDs.')
+    else:
+        selected = test if args.split == 'test' else [q for q in questions if q['id'] in development]
+    if not selected: parser.error('Empty selected split.')
+    if args.split == 'test' and (manifest.get('evaluation_kind') != 'independent_holdout'
+                                or development & {q['id'] for q in selected}):
+        parser.error('Fresh TEST requires independent freeze and zero DEV overlap.')
+    if not development <= {q['id'] for q in questions if q['split'] == 'dev'}:
+        parser.error('Generation DEV must be declared inside the dataset DEV split.')
     prompt_path = HERE / 'graph_answer_prompt.txt'; prompt = prompt_path.read_text()
     args.output_dir.mkdir(parents=True)
     metadata = run_metadata(dataset); settings = get_neo4j_settings(); contexts = {}
@@ -141,9 +151,9 @@ def main():
             result=search_graph_with_session(session,case['question'],**manifest['parameters'])
             result['index_version']=metadata['index_version']
             contexts[case['id']]=graph_context(result)
-    code_files = [Path(__file__), prompt_path, HERE/'answer_format.py',HERE/'context_policy.py',HERE/'checked_generation.py', Path('tools/rag/evaluation/rag_runner.py')]
-    hashes = {str(p): sha256_file(p) for p in code_files}
-    if args.split=='regression':
+    code_files = [Path(__file__), prompt_path, HERE/'answer_format.py',HERE/'context_policy.py',HERE/'checked_generation.py', HERE/'render_answer.py', Path('tools/rag/evaluation/rag_runner.py')]
+    hashes = {str(p.resolve().relative_to(HERE.parents[1])): sha256_file(p) for p in code_files}
+    if args.split in {'regression', 'test'}:
         dev=json.loads(args.dev_results.read_text())
         if (dev['configuration']['selected_split']!='dev'
             or set(dev['configuration']['selected_ids'])!=development
@@ -156,25 +166,29 @@ def main():
             or dev['configuration']['model']!=args.model
             or dev['run_metadata'].get('index_version')!=metadata['index_version']):
             raise RuntimeError('Generation code, prompt, data or complete DEV differs; rerun DEV before final freeze.')
-    declaration = {'method': ('Previously exposed LLM questions; regression, not independent TEST.'
-                    if manifest.get('evaluation_kind')=='historical_regression' else
-                    'Fresh LLM-generation holdout over previously measured, frozen graph queries.'),
+    declaration = {'method': ('Known-question development selection.' if args.split=='dev' else
+                    'Previously exposed LLM questions; regression, not independent TEST.'
+                    if args.split=='regression' else
+                    'Fresh query-group holdout over known sources and templates; zero DEV query-group overlap.'),
         'graph_freeze_sha256': sha256_file(frozen), 'audit_sha256': sha256_file(audit),
         'generation_files': hashes, 'model': args.model, 'temperature': 0, 'workers': args.workers,
         'dev_ids': sorted(development), 'test_ids': [q['id'] for q in test],
         'dev_count': len(development), 'test_count': len(test), 'target_accuracy': 0.92,
+        'evaluation_kind': 'historical_regression' if args.split=='regression' else 'independent_query_group_holdout' if args.split=='test' else 'development',
+        'rendering': 'Verified structured claims rendered deterministically; raw model outputs preserved.',
         'selected_split': args.split, 'selected_ids': [q['id'] for q in selected], 'selected_count':len(selected),
         'selected_dev_overlap': len(development & {q['id'] for q in selected}),
         'corrective_request_limit':1,'format_repair_limit':1,
         'repair_policy':'At most one total corrective request for format, citations or retrieval-context violations. Same question/context, no gold; retain all attempts.',
         'dev_results_sha256':sha256_file(args.dev_results) if args.dev_results else None,
         'policy': ('Evaluate every declared regression case, including DEV overlap; retain all failures and stop tuning after this freeze.'
-                   if manifest.get('evaluation_kind')=='historical_regression' else
-                   'Evaluate all 300; retain every failure; do not tune on this TEST.'),
+                   if args.split=='regression' else
+                   'Evaluate every declared fresh TEST case once; retain every failure; do not tune on this TEST.'),
         'selected_categories': dict(Counter(q['category'] for q in selected)),
-        'limitations': [('All questions and previous model outputs are known; this is historical regression.'
-                         if manifest.get('evaluation_kind')=='historical_regression' else
-                         'Known corpus/profiles and templates; model outputs for these 300 questions were not previously observed.'),
+        'limitations': [('Known development questions; not a holdout.' if args.split=='dev' else
+                         'All questions and previous model outputs are known; this is historical regression.'
+                         if args.split=='regression' else
+                         'Known corpus/profiles and templates; these query groups were not previously evaluated.'),
             'Harness supplies graph context; native OpenWebUI tool choice and browser UI are outside this experiment.',
             'Checks structured factual claims, their source support/citations and required prose values; not a general semantic judge of arbitrary prose.']}
     (args.output_dir / 'frozen_answers.json').write_text(json.dumps(declaration, ensure_ascii=False, indent=2) + '\n')
@@ -193,8 +207,9 @@ def main():
         # Grading occurs after all generation/correction decisions, never inside
         # the generator. Gold cannot influence whether another request is sent.
         for attempt in row['attempts']:
-            attempt['strict']=(grade_answer(case,attempt['answer'],oracle,context=context)['strict']
-                               if attempt.get('answer') else False)
+            candidate = attempt.get('rendered_answer', attempt.get('answer'))
+            attempt['strict']=(grade_answer(case,candidate,oracle,context=context)['strict']
+                               if candidate else False)
         if row.get('answer'):
             row.update(grade_answer(case,row['answer'],oracle,context=context))
         row['strict']=bool(row['strict'] and row['accepted_for_delivery'])
@@ -219,7 +234,7 @@ def main():
     with create_neo4j_driver(settings) as driver, driver.session(database=settings.database) as session:
         verify_freeze(manifest, run_metadata(dataset), graph_fingerprint(session), audit=audit, **manifest['parameters'])
         assert graph_version(session)==metadata['index_version'],'Index version changed during generation.'
-    assert hashes == {str(p): sha256_file(p) for p in code_files}, 'Generation code changed during TEST.'
+    assert hashes == {str(p.resolve().relative_to(HERE.parents[1])): sha256_file(p) for p in code_files}, 'Generation code changed during TEST.'
     output['by_category'] = {category: {'questions': len(rows), 'correct': sum(r['strict'] for r in rows),
         'accuracy': sum(r['strict'] for r in rows) / len(rows)} for category in sorted({q['category'] for q in selected})
         for rows in [[r for r in output['questions'] if r['category'] == category]]}

tools/answer_evaluation/graph_answer_prompt.txt

@@ -1,16 +1,11 @@
-Odpovedaj po slovensky iba z dodaných overených grafových faktov. Žiadne ďalšie vedomosti.
-Vráť JSON: {"answer": "stručná prirodzená odpoveď so zdrojmi", "status": "ok alebo no_answer", "claims": [], "source_urls": [], "count": null}.
-Každá claim obsahuje person a iba relevantné pravdivé polia: title, year, start_year, keyword, category, document_author. Nepovinné polia vynechaj; nevypĺňaj ich null ani odhadom.
-Claims musia pokryť všetky požadované fakty. Podmienky patria tej istej osobe a názov aj rok tej istej práci. Metadata category a document_author spájaj len pri rovnakom document_path.
-Presne zachovaj mená, názvy, tagy a kódy kategórií. Mená uvádzaj v pôvodnom tvare aj v texte answer. V answer uveď všetky požadované hodnoty; nezamieňaj začiatok štúdia s rokom práce.
-Title znamená výhradne názov akademickej práce alebo projektu; meno profilu nikdy nie je názov práce. Category je iba skutočný kód taxonomy.category, napríklad dp2026. Študent nie je kategória. Typ akademickej práce nie je kategória. Kódy kategórií nerozpisuj ani z nich neodhaduj rok práce. bp = bakalárska práca, dp = diplomová práca, phd = dizertačná práca; diplomová práca nie je dizertačná.
-Pri výbere kandidátov a počte uvádzaj len zodpovedajúcich kandidátov. Použi vypočítaný count, nikdy počet uvedených zdrojov. Pri porovnaní ľudí daj každému samostatnú claim.
-Source_urls musí obsahovať konkrétne relevantné URL zo zdrojov; neodhaduj ich a neupravuj. Cituj ich aj v answer. Uveď všetky zdroje potrebné na doloženie uvedených faktov.
-Ak status kontextu je no_answer, oznám, že požadované údaje sa nenašli, status no_answer, claims prázdne. Nevymýšľaj žiadny názov ani pozitívny fakt.
-Kľúčové slovo sa v tomto nástroji viaže na osobu cez jej vlastné aj dcérske dokumenty. Pri otázke na názov a kľúčové slovo spoj title, year a keyword do jednej claim pre tú istú person, aj keď majú odlišné document_path. Nevyžaduj rovnaký dokument pre tag a názov. Rovnaký dokument je povinný pri spoločnej kategórii a autorovi dokumentu.
-Pred odoslaním pri otázke na názov a tag over, že claim s názvom obsahuje súčasne person, title, year aj keyword. Nerozdeľ túto požadovanú väzbu na samostatnú claim s názvom a samostatnú claim s tagom.
-Pri rokoch porovnaj len akademické záznamy dodané v kontexte, nie poznámky o príprave. Pri počte nemiešaj nevyhovujúcich kandidátov do claims.
-Vráť iba jediný platný JSON objekt bez úvodu, vysvetlení mimo JSON a bez Markdown ohraničenia. Reťazce ohraničuj ASCII dvojitými úvodzovkami. Vo vnútri answer názvy neobaľuj ďalšími ASCII dvojitými úvodzovkami; ak ich použiješ, musia byť escapované. Roky a count sú celé čísla. Nevypisuj čiarku za poslednou položkou.
-Rešpektuj complete, truncated, truncation_reasons a index_version. Ak truncated=true alebo complete=false a uvádzaš výsledky, do answer vlož vetu „Výpis je neúplný.“ Neoznačuj výpis za úplný zoznam ani za všetky existujúce práce. Pri complete=null úplnosť nemožno zaručiť. Ak aggregation.complete=true, počet môže byť presný aj pri orezaných dôkazoch; vtedy uveď „Počet je úplný, dôkazy sú orezané.“ Inak z počtu zobrazených položiek neodvodzuj celkový počet.
-Všetky URL v answer aj source_urls musia byť presne zo sources. Typ práce (napríklad diplomový projekt) nie je kľúčové slovo; keyword uveď iba ak ho dokladá fakt keyword pre tú istú person.
-answer_candidates sú doložené spojenia požadovaného názvu alebo roku a tagu cez tú istú osobu, nie nové odhady. Ak je taký kandidát poskytnutý a status kontextu je ok, odpovedz podľa neho a neodmietaj ho len preto, že tag pochádza z dcérskeho dokumentu. V claims ponechaj jeho person, year, keyword a prípadne title spolu v jednom objekte a cituj všetky jeho podklady. Aj pri samotnej kombinácii tag + rok práce sa tag viaže na osobu, nemusí byť v dokumente obsahujúcom rok práce.
+Vyber iba pravdivé štruktúrované tvrdenia z dodaného graph_context. Nepoužívaj vlastné vedomosti, zdrojový oracle ani údaje z otázky ako dôkaz.
+Vráť jediný platný JSON objekt: {"answer":"","status":"ok","claims":[],"source_urls":[],"count":null}. Answer MUSÍ zostať prázdny reťazec. Prirodzený text, citácie a upozornenie na úplnosť zostaví aplikácia z overených polí. Nepíš voľný text ani Markdown mimo JSON. Všetky reťazce ohraničuj ASCII dvojitými úvodzovkami; úvodzovky v hodnotách escapuj.
+Každá claim obsahuje person a len požadované polia, ktoré sú priamo doložené faktami: title, year, start_year, keyword, category, document_author. Nepovinné polia vynechaj. Roky a count sú celé čísla. Presne zachovaj mená, názvy, tagy a kódy kategórií. Nevytváraj nové osoby, hodnoty ani typy tvrdení.
+Title je názov akademickej práce/projektu, nikdy názov profilovej stránky alebo generický odkaz na rešerš. Year je rok tej istej práce ako title. Start_year je explicitne doložený začiatok štúdia; rok adresára nie je dôkaz. Keyword je skutočný tag osoby, nikdy typ práce ako diplomový projekt. Category je presný taxonomy.category kód. Document_author je autor dokumentu, nie automaticky študent.
+Všetky podmienky jednej claim patria tej istej osobe. Title a year patria rovnakému akademickému záznamu. Category a document_author musia byť doložené tým istým document_path.
+Tag patrí osobe cez jej profil aj dcérske dokumenty. Pre kombináciu názvu/roku a tagu nevyžaduj rovnaký dokument. Answer_candidates sú už doložené spojenia cez tú istú osobu; zachovaj ich person, year, keyword a prípadne title spolu v jednej claim. Neodmietaj toto spojenie len pre odlišné document_path.
+Claims musia pokryť všetky relevantné výsledky dotazu a požadované väzby. Pri porovnaní ľudí uveď každého v samostatnej claim. Pri výbere kandidátov vynechaj nevyhovujúce osoby. Pri názve práce uveď aj rok, ak je vo faktoch známy.
+Source_urls musí obsahovať konkrétne relevantné source_url zo sources pre všetky tvrdenia. Nevymýšľaj ani neupravuj URL. Ak názov/rok dokladá profil a tag dcérsky dokument, cituj oba potrebné zdroje. Nepoužívaj URL iných osôb ako náhradnú citáciu.
+Count môže byť číslo iba pri aggregation.complete=true a musí sa rovnať aggregation.count. Inak count=null, aj pri no_answer. Počet nie je počet zobrazených zdrojov. Presný agregovaný počet môže byť úplný aj pri orezanom výpise dôkazov.
+Pri status kontextu no_answer vráť status=no_answer, claims=[], count=null. Pri ambiguous alebo error zachovaj tento status a prázdne claims. Nevymýšľaj pozitívnu odpoveď. Nulová úplná agregácia so status=ok je platný výsledok s count=0.
+Complete, truncated, truncation_reasons a index_version nemenia pravdivosť faktov. Význam a obmedzenia výsledku oznámi aplikácia. Answer nechaj prázdne aj pri neúplnosti, nenájdení a chybe.

tools/answer_evaluation/render_answer.py

@@ -0,0 +1,29 @@
+"""Render only validated structured facts; never infer facts or read gold."""
+from copy import deepcopy
+
+
+def render_graph_answer(answer: dict, context: dict) -> dict:
+    out = deepcopy(answer)
+    lines = []
+    if out['status'] == 'ok':
+        labels = {'title': 'názov práce', 'year': 'rok práce', 'start_year': 'začiatok štúdia',
+                  'keyword': 'kľúčové slovo', 'category': 'kategória', 'document_author': 'autor dokumentu'}
+        for claim in out['claims']:
+            values = [f'{label}: {claim[key]}' for key, label in labels.items() if claim.get(key) is not None]
+            lines.append(claim['person'] + (' — ' + '; '.join(values) if values else '') + '.')
+        if out.get('count') is not None:
+            lines.append(f'Počet vyhovujúcich osôb: {out["count"]}.')
+        if context.get('complete') is False or context.get('truncated'):
+            lines.append('Výpis je neúplný.')
+        elif context.get('complete') is None:
+            lines.append('Úplnosť vyhľadania nemožno zaručiť.')
+    elif out['status'] == 'ambiguous':
+        lines.append('Požiadavka je nejednoznačná; údaje nemožno spoľahlivo určiť.')
+    elif out['status'] == 'error':
+        lines.append('Vyhľadanie zlyhalo; odpoveď nemožno pripraviť.')
+    else:
+        lines.append('Požadované údaje sa v dostupných podkladoch nenašli.')
+    if out['source_urls']:
+        lines.append('Zdroje: ' + ', '.join(out['source_urls']))
+    out['answer'] = '\n'.join(lines)
+    return out

tools/graphrag/evaluation/README.md

@@ -1,6 +1,22 @@
 # GraphRAG evaluácia
 
-Aktuálna verzia je **v6**, tag `graphrag-v6-frozen`.
+Aktuálna verzia je **v7**, tag `graphrag-v7-frozen`.
+[Správa v7](results/final_v7_2026-10-26/REPORT.md): graf DEV 1 046/1 046,
+nový query-group TEST 80/80, model DEV 46/46, finálne odpovede TEST 80/80
+(prvý pokus 77/80). Všetkých 591 testov prešlo. Stará regresia má 319/320;
+jeden nepodložený modelový názov sa nedoručil. TEST sa neprekrýva so žiadnou odhalenou query skupinou;
+korpus, ľudia a šablóny sú známe. Dve kategórie majú len 2 a 7 nových skupín.
+
+CLI vyžaduje explicitné `--questions`, `--audit`, `--freeze-manifest`
+(pri TESTe) a nový výstupný adresár. Použite `graphrag_questions_v7.json`,
+`results/v7_preparation/source_audit.json`, `frozen_v7.json`.
+Generátor `benchmark_v7` vytvára otázky jednorazovo z raw zdrojov, bez
+predikcií. Odstránil sa odhad začiatku štúdia z adresára, nie nedokumentované
+otázky z merania. [Príkazy a postup](../../../docs/graphrag_v7_2026-10-26_worklog.md).
+
+## Historická v6
+
+Historická verzia je **v6**, tag `graphrag-v6-frozen`.
 [Finálna správa v6](results/final_v6_2026-10-25/REPORT.md) oddeľuje grafový
 DEV 722/722, regresiu 320/320 a generované odpovede modelu.
 Generovanie má strict **316/320 (98,75 %)**; všetkých **565 testov** prešlo.

tools/graphrag/evaluation/benchmark_v2.py

@@ -99,7 +99,8 @@ def binding_source_evidence(binding: dict, oracle: dict) -> list[dict]:
                              'year_basis': row['year_basis']})
     if 'start_year' in binding:
         evidence.append({'document_path': person['profile'], 'sha256': oracle['documents'][person['profile']]['sha256'],
-                         'field': 'start_year', 'value': binding['start_year']})
+                         'field': 'start_year', 'value': binding['start_year'],
+                         'basis': person.get('start_year_basis', 'explicit_text'), 'confidence': 'documented'})
     return evidence
 
 
@@ -378,6 +379,9 @@ def validate_case(case: dict, oracle: dict) -> list[str]:
         if constraints.get('unknown_person'):
             if norm(constraints['unknown_person']) in {norm(name) for name in oracle['people']}:
                 issues.append('negative_person_actually_exists')
+        elif constraints.get('missing_start_year'):
+            if names and all(oracle['people'][name]['start_year'] is not None for name in names):
+                issues.append('negative_start_year_actually_documented')
         elif constraints.get('shared_academic_year'):
             year_sets = [{r['year'] for r in academics(oracle, name) if r['year'] is not None} for name in names]
             if year_sets and set.intersection(*year_sets):
@@ -408,7 +412,7 @@ def validate_case(case: dict, oracle: dict) -> list[str]:
         if names and set(graph['expected_entity_set']) != set(oracle['people'][names[0]]['categories']):
             issues.append('incomplete_category_set')
     if graph.get('expected_count') is not None:
-        years = re.findall(r'\b20\d{2}\b', case['question'])
+        years = re.findall(r'\b\d{4}\b', case['question'])
         if not years:
             issues.append('count_without_supported_year_constraint')
         else:

tools/graphrag/evaluation/benchmark_v4.py

@@ -33,7 +33,7 @@ def logical_group(case: dict) -> str:
     names = sorted(norm(row['value']) for row in graph['start_entities'] if row['type'] == 'person')
     bindings = graph.get('expected_bindings', [])
     keywords = sorted({norm(row['keyword']) for row in bindings if row.get('keyword')})
-    years = sorted({int(value) for value in re.findall(r'\b20\d{2}\b', case['question'])})
+    years = sorted({int(value) for value in re.findall(r'\b\d{4}\b', case['question'])})
     # A shared-year query has no requested year; its complete answer is derived.
     selectors = {'keywords': keywords, 'requested_years': years}
     if category == 'document_author_category_multi_hop':

tools/graphrag/evaluation/benchmark_v7.py

@@ -0,0 +1,91 @@
+"""Declare a fresh query-group holdout using raw sources only, before retrieval.
+
+All 1,042 previously exposed questions become DEV. No profile is new; this is
+a new-combination holdout over a known corpus, not an unseen-person experiment.
+"""
+import copy
+import json
+import random
+from collections import Counter
+
+from .benchmark_v2 import BASE, audit_dataset, make_case, validate_case, write_json, binding_source_evidence
+from .benchmark_v4 import build_pool, logical_group, refresh_titles
+from .source_oracle import load_source_oracle, norm
+
+QUOTAS = {'shared_keyword_pair': 10, 'same_academic_year_pair': 10,
+          'same_start_year_pair': 10, 'person_title_year': 10, 'person_keyword_title': 10,
+          'document_author_category_multi_hop': 2, 'person_multiple_categories': 7,
+          'candidate_set_aggregation': 7, 'conjunctive_disambiguation': 6, 'negative_lookup': 8}
+SEED = 20261026
+
+
+def main():
+    output = BASE/'graphrag_questions_v7.json'; artifacts = BASE/'results/v7_preparation'
+    if output.exists(): raise RuntimeError('Refusing to overwrite declared questions.')
+    oracle = load_source_oracle()
+    old = json.loads((BASE/'graphrag_questions_v6.json').read_text())
+    dev = []; changes = []
+    exposed = {logical_group(q) for q in old}
+    for original in old:
+        case = copy.deepcopy(original); case['split'] = 'dev'
+        names = [r['value'] for r in case['graph']['start_entities'] if r['type'] == 'person']
+        if any(b.get('start_year') is not None and oracle['people'][b['person']]['start_year'] is None
+               for b in case['graph']['expected_bindings']):
+            case = make_case(case['id'], 'dev', case['category'], case['question'], names, oracle, [],
+                no_answer=True, reason='The original Markdown does not explicitly document every requested study-start year.',
+                negative_constraints={'missing_start_year': True})
+            changes.append({'id': case['id'], 'old_bindings': original['graph']['expected_bindings'],
+                            'new_status': 'no_answer', 'reason': 'Directory cohort is not a documented study-start year.'})
+        elif case['should_answer']:
+            case['source_evidence'] = [e for b in case['graph']['expected_bindings'] for e in binding_source_evidence(b, oracle)]
+        case['graph']['partition_group'] = logical_group(case); dev.append(case)
+    # Fixed audit regressions; these are explicitly DEV, not extra easy TEST negatives.
+    for year in [1999, 1899, 2100, 9999]:
+        case = make_case(f'v7-dev-year-{year}', 'dev', 'negative_lookup',
+            f'Aký je názov práce osoby Ján Holp v roku {year}?', ['Ján Holp'], oracle, [],
+            no_answer=True, negative_constraints={'year': year})
+        case['graph']['partition_group'] = logical_group(case); dev.append(case)
+    exposed.update(logical_group(q) for q in dev)
+    texts = {norm(q['question']) for q in dev}; rng = random.Random(SEED)
+    holdout = []; capacity = {}
+    for category, pool in build_pool(oracle).items():
+        unique = {}
+        for candidate in pool:
+            case = refresh_titles(candidate, oracle); group = logical_group(case)
+            if group not in exposed and norm(case['question']) not in texts:
+                unique[group] = case
+        capacity[category] = len(unique)
+        candidates = [unique[key] for key in sorted(unique)]; rng.shuffle(candidates)
+        quota = QUOTAS[category]
+        if len(candidates) < quota: raise RuntimeError(f'Insufficient source capacity: {category}')
+        for case in candidates[:quota]:
+            case['id'] = f'fresh7-{len(holdout)+1:03d}'; case['split'] = 'test'
+            case['graph']['partition_group'] = logical_group(case); holdout.append(case)
+    questions = dev + holdout
+    for case in questions:
+        issues = validate_case(case, oracle)
+        if issues: raise RuntimeError(f'{case["id"]}: {issues}')
+    write_json(output, questions)
+    audit = audit_dataset(questions, oracle, output)
+    audit['partition_policy'] = {'method': 'disjoint logical query groups',
+        'shared_source_profiles': True, 'seed': SEED, 'previously_exposed_groups': sorted(exposed),
+        'limitations': ['Known corpus, persons and template families; only the query/fact combinations are fresh.',
+                        'Two remaining document-author/category groups and seven category-list groups constrain category coverage.',
+                        'AI source annotation, not human annotation. No tuning on this TEST.']}
+    write_json(artifacts/'source_audit.json', audit)
+    write_json(artifacts/'source_corrections.json', {'questions': changes,
+        'policy': 'Only explicit study-start text is a fact. Cohort remains separate metadata; missing means unknown.'})
+    write_json(artifacts/'regression_ids.json', [q['id'] for q in old if q['split'] == 'test'])
+    ids = set(json.loads((BASE/'results/v6_preparation/generation_dev_ids.json').read_text()))
+    ids.update(f'v7-dev-year-{year}' for year in [1999, 1899, 2100, 9999])
+    write_json(artifacts/'generation_dev_ids.json', sorted(ids))
+    write_json(artifacts/'protocol.json', {'seed': SEED, 'dev_count': len(dev), 'fresh_test_count': len(holdout),
+        'test_categories': dict(Counter(q['category'] for q in holdout)), 'test_positives': sum(q['should_answer'] for q in holdout),
+        'test_negatives': sum(not q['should_answer'] for q in holdout), 'available_groups': capacity,
+        'old_gold_corrections': len(changes), 'regression_count': 320, 'llm_dev_count': len(ids),
+        'target': .92, 'policy': 'Freeze after complete DEV; run fresh TEST once, retain every failure. Regression remains separate.',
+        'annotation': 'Independent raw-source oracle plus AI source review; not human review.'})
+    print(json.dumps({'dev': len(dev), 'test': len(holdout), 'source_corrections': len(changes), 'quotas': QUOTAS}))
+
+
+if __name__ == '__main__': main()

tools/graphrag/evaluation/check_openwebui.py

@@ -88,7 +88,7 @@ def main():
     from tools.graphrag.build_graphrag import graph_version
 
     parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument('--questions', type=Path, default=BASE / 'graphrag_questions_v4.json')
+    parser.add_argument('--questions', type=Path, required=True)
     parser.add_argument('--freeze-manifest', type=Path, required=True)
     parser.add_argument('--audit', type=Path, required=True)
     parser.add_argument('--output', type=Path, required=True)

tools/graphrag/evaluation/evaluate_graphrag.py

@@ -54,13 +54,13 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument(
         "--questions",
         type=Path,
-        default=QUESTIONS_PATH,
+        required=True,
     )
 
     parser.add_argument(
         "--output-dir",
         type=Path,
-        default=RESULTS_DIR / 'development_v4',
+        required=True,
     )
 
     parser.add_argument(

tools/graphrag/evaluation/freeze_graphrag.py

@@ -88,7 +88,7 @@ def make_freeze(dev: dict, questions: list[dict], fingerprints: dict, audit: dic
 def main() -> None:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument('--dev-results', type=Path, required=True)
-    parser.add_argument('--questions', type=Path, default=Path(__file__).parent / 'graphrag_questions_v4.json')
+    parser.add_argument('--questions', type=Path, required=True)
     parser.add_argument('--audit', type=Path, required=True)
     parser.add_argument('--output', type=Path, required=True)
     parser.add_argument('--purpose', choices=['independent_holdout','historical_regression'],

tools/graphrag/evaluation/frozen_release_v7.json

@@ -0,0 +1,137 @@
+{
+  "release": "graphrag-v7-frozen",
+  "evaluation_kind": "fresh_query_group_holdout_plus_historical_regression",
+  "graph_freeze_sha256": "9e97794200520d0f2288e452c9e186c0291a54c63d2c1382f693a9753e4081cf",
+  "runtime_files": {
+    "app/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "app/main.py": "ff1adeff8b770047139fe3f632ad837517c95ec04649849c4315d6f892f9eef3",
+    "app/routes.py": "dfb82171fea33d33b6e48cb238d2c19a29254b47a9a87308f4864cdbac53fa7b",
+    "app/security.py": "24f437aff6642477049f02fdccfb3429c15e610d044bbb66128903771510055f",
+    "requirements.txt": "26a6d097727eba03c1b6c71cd02b06373ef01ce03e567dd2bc4357ae330e5a1a",
+    "tools/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/answer_evaluation/__init__.py": "795be75bf60e4aa9209e5b820207e52e8420c41a4dc2e37d0b6018cd2934a51a",
+    "tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
+    "tools/answer_evaluation/check_context_scenarios.py": "99dbae2747183622e0823cd8f6536f4ff38ef56d0e41dd7855096d47af3b1c37",
+    "tools/answer_evaluation/check_deployment.py": "14618b102c0ade954dc73cbc5036a628166f0eb703d435b679d1a000d045ab5a",
+    "tools/answer_evaluation/check_rag_runtime.py": "1e1b8bd48ac4d668ec057fee162ef30d34ae20ae026361bc8cc296bc455d4067",
+    "tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
+    "tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
+    "tools/answer_evaluation/evaluate_graph_answers.py": "d81504e8fe51598c6bd1b0d482f61b9b6f51f87ba3f38b03ae1bc584675a3b76",
+    "tools/answer_evaluation/graph_answer_prompt.txt": "fc9f2f8d52e70a8d67b9f4814b3f15e6be64bf88afe53ce4db4f037bde00e064",
+    "tools/answer_evaluation/rag_delivery.py": "b83f93d446f44fdd7d41c12eeeaae69c0b02908076624f4eb3f4f72b9b88dbc8",
+    "tools/answer_evaluation/regrade_citations.py": "330f437aec12927fd8c99b0aa601373f6ab26cb111971a6fb06098931d1e4aab",
+    "tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002",
+    "tools/graphrag/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/graphrag/build_graphrag.py": "256292101d6d7d1946e926e2717ac45465b169f4091c51ecfed4f1a779dfd700",
+    "tools/graphrag/evaluation/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/graphrag/evaluation/audit_benchmark.py": "6619c818a52fc284e87ed6c48a503a54a3c9dd95447463ed78f075b7c58bad80",
+    "tools/graphrag/evaluation/benchmark_v2.py": "62a0b6af852d5ae8060c06db3f983ef4b797d74443f9df25a6b8bd77bf133330",
+    "tools/graphrag/evaluation/benchmark_v3.py": "30bae6c605ed26f4b49973acec799087c4e2b97a883ccedcafb965dd909c4c4d",
+    "tools/graphrag/evaluation/benchmark_v4.py": "7a731fe649d3d5170045b8bbcabf31ad527e5d26a934d434a4be3a4c94dc5c38",
+    "tools/graphrag/evaluation/benchmark_v6.py": "f955e600d17358fafcb33bab772aef09d787452f6b33352e299fdd42669aaf34",
+    "tools/graphrag/evaluation/benchmark_v7.py": "9477cd6e9a6dd0558837cf044535de75a8f101b17115e28e57a82d96d1aa01cb",
+    "tools/graphrag/evaluation/check_openwebui.py": "863801048231d474ba82930cf6bf6853b4410f8295b3f97fd3d5934852431fbe",
+    "tools/graphrag/evaluation/evaluate_graphrag.py": "22f48e9ac6dfae92d79ad648ca173547c6a37f98aa8199f7c37595f1b0b186f9",
+    "tools/graphrag/evaluation/evidence_contract.py": "78e24264c176e163e1782fef1634b6c45c7f31bd4692726d02a30ab45a06648e",
+    "tools/graphrag/evaluation/freeze_graphrag.py": "d8764d1aeb72208de55604785acff880002da1f818d8a8ad9477cf39e1f29efb",
+    "tools/graphrag/evaluation/graphrag_evaluator.py": "7f99bf0f97974058650b619ea49a8ee32284cb4b324f1ffab9baf54f5c4495f4",
+    "tools/graphrag/evaluation/reproducibility.py": "0e2dab6ffb17691f05f9d81dbec12ffc0f2e4e0b1a2114a75b66a5165cf7b4d2",
+    "tools/graphrag/evaluation/source_oracle.py": "aed0309435fd21bc274fde7905a0720f68f359b3c30bb0b912e3371ebf0d8072",
+    "tools/graphrag/evaluation/source_validation.py": "9671b8978623888dbed8b8b2d1b6b01c1ce73cb0a512b03919e48873990f15ed",
+    "tools/graphrag/graph_db.py": "c4b1a2edb765856b807398e4c243a8d17753529a5d2dfdb51791e28bc8de17f7",
+    "tools/graphrag/graph_search.py": "c41cbfbac0d5daa1d0b51f5f4b798a15688bbd2e0498d2cbe93c6584a99f6b7d",
+    "tools/graphrag/graphrag.py": "dbc75403aa038e4021b5a70d1f17c62784b478252a94bac8db682bf38dcb39b3",
+    "tools/graphrag/neo4j_smoke.py": "fb5244c72c29ef2acf96d1e7f54335ff168eed4119db1692025eb1b99194b7bb",
+    "tools/graphrag/publication.py": "fba1fa0d0bed325fd87f8d3593baae9271de1766b772d7214c7f63fd3f1c37b6",
+    "tools/graphrag/tests/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/graphrag/tests/test_evaluation_pipeline_v2.py": "63d1b3e80ae3c7b563374fd7ab4aa5ffad3c614eafbd15142c577abd7f756bf9",
+    "tools/graphrag/tests/test_evaluation_v5.py": "9387e9d1418ebab87530986b0e29dac31cf33d762777a29a31cff24cc72e552a",
+    "tools/graphrag/tests/test_evaluation_wide.py": "1b28bb74fd6b9457c869186b88606f1f7863d573aa3ab5226162e878f236d99e",
+    "tools/graphrag/tests/test_evaluator_connected_v2.py": "29d7c9838ed1ff37ea3276df6feb1b6142158c43a0525712f490b648181882de",
+    "tools/graphrag/tests/test_graph_publication.py": "96e53c1ae28e36c7a4290e1e60c8b398bf095ceaf883e819e1a477e2672b11c0",
+    "tools/graphrag/tests/test_graph_rebuild_atomic.py": "fe37d4fff0b1f0fb1d565f0a96152b9b4e336837bc78b5977cf3b4879d7fbaf3",
+    "tools/graphrag/tests/test_graph_search.py": "eb58661aef9dafe6bf7188fcdc2d79cef6652d57c64d1b81ae81a52e8e0204d7",
+    "tools/graphrag/tests/test_graph_search_entity_v2.py": "0ef1a8ebd513c5a4507eb38fcf5ae6137c69e9901b5a2975ba513be5902ac693",
+    "tools/graphrag/tests/test_graph_search_neo4j.py": "fec2bf72de8f752f17c9164fee5754bdb2e153828855e47efbf0000dbe82cf3b",
+    "tools/graphrag/tests/test_graph_search_regressions.py": "9db7779aeb8ed9e8c0f92c1657011a81080aff3c928649e811358e3009a18c48",
+    "tools/graphrag/tests/test_graphrag.py": "9ee8629d6b528dd78c2ddfc25cee95e17613f45b454fc0a38d4c591862637a9f",
+    "tools/graphrag/tests/test_graphrag_evaluator.py": "1c6d08aeec9ae56ac4123239a5aed12a0081d66cc3f34b2047d95764579d7298",
+    "tools/graphrag/tests/test_graphrag_freeze.py": "c4b760baf4c7ebf13b16bd4cd3c51874f5512eef3b4c61470dc4d428b0303d9f",
+    "tools/graphrag/tests/test_graphrag_research_v2.py": "711008a91ab8a58ae0854e7f90329bcd42e5189fde2e2001589e97b0c22c5198",
+    "tools/graphrag/tests/test_graphrag_root_title.py": "085651fe6e6081521a5dde9554abfee93456250108cf15978df25378d886a46c",
+    "tools/graphrag/tests/test_source_oracle_v2.py": "0afdbe235e185ffe9f38aa667f90f5a6c20504fe29e9ac572642b39a8eb7faca",
+    "tools/graphrag/tests/test_title_labels.py": "2983ead68ad6e2e346d24af0fc96dc6771c163e9c2a351af426a1e0383998e6f",
+    "tools/index_state.py": "c33920a63d04009f11ef23ca9e9eee9bf86453b7d0586d027bf3b39156afea0d",
+    "tools/publication.py": "f99d445073390a733afc8d849b9fcd4b4b0890bb54d136db67d4015b78436e1a",
+    "tools/rag/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/rag/build_chunks.py": "fe060e25b138abcb689a6eb1b1e993d7017a4b3c5dd63e7cd275e1279da6053b",
+    "tools/rag/build_sqlite_index.py": "107c0987e9ec4086ff41f9e7ae30195a67be34616875635e705dc79236e79416",
+    "tools/rag/common.py": "8497c46d5c6c27a7c76df1568360ad5e9d00d5e5a7ef186e758960cc79d81d1c",
+    "tools/rag/embedding_utils.py": "a2ea0b5be1a1782efff1c85c0eabfedba043e0e8bf7eb8b6617eab60eb4559c3",
+    "tools/rag/evaluation/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/rag/evaluation/evaluate_rag_answers.py": "945c2da655537fb37fe82a08e10dc3ed66a2d90980921f61f61efb171a4a26a2",
+    "tools/rag/evaluation/evaluate_retrieval.py": "aba865751ab74d9c2b962548b1e476508dbe6cf7bf811bda79e25f853061cbd7",
+    "tools/rag/evaluation/metrics.py": "40b7d16f7729599c1d2fa2370f7cb40dfa04a968cdadef3c4fae83a168202c28",
+    "tools/rag/evaluation/rag_answer_data.py": "9a369e9ced254cf723a5d3006877959c6f144f4c770e382835af776f7f569762",
+    "tools/rag/evaluation/rag_answer_state.py": "da4ab19f22378899d5c13e93e803a978518f8edcdd62e4990a06aae12d8eb469",
+    "tools/rag/evaluation/rag_metrics.py": "3e95922df90dd1b757a550040a2d9bd4f857e17d179f60170d7cc881a08f94b6",
+    "tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1",
+    "tools/rag/evaluation/retrieval_runner.py": "fe3dd05459f825d72252b7a0841f28004fc81537096a7de487807e4d5f0c4c45",
+    "tools/rag/rag_document_expansion.py": "d565b9794e30bc0f387c791a0557d5d03f2c1fa2fe5db53e49a9b19b1158f48e",
+    "tools/rag/rag_query_evidence.py": "90fced0fecd8ef976a1537506d2e7e8e9d3ba22a6371131f978fb91df5cc0b19",
+    "tools/rag/rag_supplemental_expansion.py": "a593579e543149633e3146c3886536a9196aff43f4df6f3d7ade94ca16dd243a",
+    "tools/rag/rag_utils.py": "4f898cf0facc2ae9f2ea28a36f48d4ed37bd7fc716e097b319bcf4d36a4a4f03",
+    "tools/rag/rebuild_index.py": "6a4d86a551d1141c403a0f6570c568896af37faeaa8c67cac170d2e6a393ca06",
+    "tools/rag/scan_zpwiki.py": "ef770ce4edb6352ff0690af5783766c95062a0e56f638cca6e3d929973728273",
+    "tools/rag/search_core.py": "624c8ddfe3edf78d06014579b747a59b2e33c0191a7cf90cdfd624b52ed5e96b",
+    "tools/rag/search_db.py": "cc63aeaf2854e87b7448c3978959569b097074cdb16a9b2c16b2c37a4677fdbb",
+    "tools/rag/search_utils.py": "69597bf5785f7852b1131a0648d9d2c65b6babba9ad279581f08b40d185ee646",
+    "tools/rag/tests/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+    "tools/rag/tests/conftest.py": "0c94402d5ca971233d9802ded81cdb39ae7dfddda5eb9126e2035f7599a8f1c7",
+    "tools/rag/tests/test_chunking.py": "94904a20cfed0512d508017eaadbb43ef45ca74a5e96631241f6ca481222c498",
+    "tools/rag/tests/test_common.py": "86f0aecb79b7a3764112c833ea3b732073c073a5d570ef8c99900fb2a3f04430",
+    "tools/rag/tests/test_database.py": "31fc3553d76127c2d7e2c113b9053e537a72bb92f57a9c24cad6a579a1ffc589",
+    "tools/rag/tests/test_e5_1_remaining.py": "a8a52d5204e7376f33bb1aef96aad20990126a0b92266596b7cf8dd0dcbea4bc",
+    "tools/rag/tests/test_evaluate_rag_answers.py": "2143f3c9efbacf7b9119eeb9ae69e6cd7ea3f428ff5076d8e9973d0943675296",
+    "tools/rag/tests/test_live_data.py": "3c5e04d4a29ec8360d89881be980dacacccf0e12594fe51a22ec23c32c98b3e3",
+    "tools/rag/tests/test_rag.py": "13e97c88bf4afe45a8d633b269d110890f0f646887cee861bbf55ada0be7ad24",
+    "tools/rag/tests/test_rag_document_expansion.py": "9adf39c7a9b683b1d40c36fce576272e71869b7c1236c371c1b077820874eb7e",
+    "tools/rag/tests/test_rag_metrics_morphology.py": "070ee314d326e4c7218e7c19959a94076a2b241c4f19b9ecd9ca1fb3cc2639d4",
+    "tools/rag/tests/test_rag_metrics_semantic.py": "3642394d2636edd7420589348fb75bf1d214b539983b7e620008d626deffabf0",
+    "tools/rag/tests/test_rag_query_evidence.py": "c3fb5115306171010759b0cfd0cd05d1439d6a57de049efea02330ae7c553bfc",
+    "tools/rag/tests/test_rag_query_evidence_topn.py": "ebf01ad57a52f030b908b7fb0312d8d32cf9f60ea2953c5a3fb8849c1f2dceaf",
+    "tools/rag/tests/test_rag_runner.py": "1813086999d1e9826b42c2ac7978ea77d29e1c4b12f5e606f1fba572829e24ea",
+    "tools/rag/tests/test_rag_runner_routing.py": "c9454e5dddfdb7afc013567811efec15545daf20733b7167f78f2dd9ff0380d6",
+    "tools/rag/tests/test_rag_runner_tool_prompt.py": "073403907717d8d388c28ea92d1d1f9426edab2e8bd320c8f6329bcbff082e6c",
+    "tools/rag/tests/test_rag_supplemental_expansion.py": "65449054e2d94396def0dc8d505a9573306adbf184820ddd6c9b5ea92e165f8b",
+    "tools/rag/tests/test_rebuild_index.py": "aa0688b81e4b699ce99e97a4b8c1a6d4a49c5c153caa02301ebd5a5d8be764a4",
+    "tools/rag/tests/test_scan_zpwiki.py": "324510ca504dff11ecf39b28aff26b44a358e14084b3b015c0c410690d8aa129",
+    "tools/rag/tests/test_search.py": "82c96d35719d2233e4636b6f25d89b22f004848d8bae46a9705ee384d4abe9a8",
+    "tools/retrieval_contract.py": "b8fa272c3babf924a3ce8b937fe7ae437cd959ba9f7667193ac67c18d2c5b58d",
+    "tools/sync_indexes.py": "5b909b4394c24e5ce22848594b16d9aea45fb06418f5cd92bb6aa6381a0ccf10",
+    "tools/tool_health.py": "8ad37345889ad58a737edac5960b5b6c5002af2397d60248f434e03fb4eddd89"
+  },
+  "deployment_configuration": {
+    "Dockerfile": "fc7aefe4a99222daedc321acf8359773dc24eff82be04e688507e135956a92a6",
+    "docker-compose.yml": "e33b65c58f155ac057b01be272bc323fab3ff7eb4f51c9e4ca1b53904cc17af4"
+  },
+  "measured_artifacts": {
+    "tools/graphrag/evaluation/graphrag_questions_v7.json": "0f78e9749de369d752b6e1079f174cf3f6b390f4718a672119dfd43fa5d9dfac",
+    "tools/graphrag/evaluation/results/final_v7_2026-10-26/graphrag_dev.json": "d1a5d7b6d5aff3c0002f36a60cebc07627a79ec6d57de119dd587e8ab70143b0",
+    "tools/graphrag/evaluation/results/final_v7_2026-10-26/graphrag_test.json": "07d3d73c1b7ab0805549d2d3ebea0c02d1c00eacf3d6092911e34e9787cd9612",
+    "tools/graphrag/evaluation/results/final_v7_2026-10-26/historical_graph_summary.json": "f0386fde85ed1779d1262264614d4d93faf90cca9a3671cd7591ff6b57759a9b",
+    "tools/graphrag/evaluation/results/final_v7_2026-10-26/summary.json": "cdc7809eb1033bdf27569ea80deb53b3e82cc2beb0141e64d6251756737a396c",
+    "tools/graphrag/evaluation/results/v7_preparation/source_audit.json": "7825d4f3e4e535eb5f49e975dfef363b9b813f1de5c97e2387e5c2f3e578e6a5",
+    "tools/graphrag/evaluation/results/v7_preparation/protocol.json": "7ae343bba36fe25e8603f00325fe2dec1f5753616db9ef9c54f74ccffc1a7d83",
+    "tools/graphrag/evaluation/results/v7_preparation/source_review.csv": "7ee21230e5dda7be9dc847b8ada6111d14d38f7b075a9c64d38f3ff69950b5c0",
+    "tools/graphrag/evaluation/results/answers_v7_dev_2026-10-26/answers_dev.json": "4af9b52c250c837707c0690d892ec833347b3d3ee71d338ec631f6450eef8045",
+    "tools/graphrag/evaluation/results/answers_v7_test_2026-10-26/answers_test.json": "19d5e6ef484495a0acc1402b2d8f6ba91b3b63b8229d1118cd75d763082c148a",
+    "tools/graphrag/evaluation/results/answers_v7_regression_2026-10-26/answers_regression.json": "4bad6d54955b7645d861ea1ae2b412625b1ea032297caa57984f3b0cf21e3a25"
+  },
+  "policy": "Runtime code, prompt and gold remained unchanged throughout the final TEST and historical regression. All failures retained.",
+  "secrets_excluded": [
+    ".env",
+    "credentials",
+    "private keys"
+  ]
+}

tools/graphrag/evaluation/frozen_v7.json

@@ -0,0 +1,252 @@
+{
+  "schema_version": 2,
+  "evaluation_kind": "independent_holdout",
+  "partition": {
+    "method": "disjoint logical query groups",
+    "logical_group_overlap": 0,
+    "previously_exposed_test_groups": 0,
+    "dev_group_count": 901,
+    "test_group_count": 80,
+    "dev_profile_count": 72,
+    "test_profile_count": 55,
+    "source_profile_overlap": 55,
+    "limitations": [
+      "Known corpus, persons and template families; only the query/fact combinations are fresh.",
+      "Two remaining document-author/category groups and seven category-list groups constrain category coverage.",
+      "AI source annotation, not human annotation. No tuning on this TEST."
+    ]
+  },
+  "frozen_at_utc": "2026-10-26T00:14:20.132176+00:00",
+  "fingerprints": {
+    "implementation": {
+      "tools/graphrag/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "tools/graphrag/build_graphrag.py": "256292101d6d7d1946e926e2717ac45465b169f4091c51ecfed4f1a779dfd700",
+      "tools/graphrag/evaluation/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "tools/graphrag/evaluation/audit_benchmark.py": "6619c818a52fc284e87ed6c48a503a54a3c9dd95447463ed78f075b7c58bad80",
+      "tools/graphrag/evaluation/benchmark_v2.py": "62a0b6af852d5ae8060c06db3f983ef4b797d74443f9df25a6b8bd77bf133330",
+      "tools/graphrag/evaluation/benchmark_v3.py": "30bae6c605ed26f4b49973acec799087c4e2b97a883ccedcafb965dd909c4c4d",
+      "tools/graphrag/evaluation/benchmark_v4.py": "7a731fe649d3d5170045b8bbcabf31ad527e5d26a934d434a4be3a4c94dc5c38",
+      "tools/graphrag/evaluation/benchmark_v6.py": "f955e600d17358fafcb33bab772aef09d787452f6b33352e299fdd42669aaf34",
+      "tools/graphrag/evaluation/benchmark_v7.py": "9477cd6e9a6dd0558837cf044535de75a8f101b17115e28e57a82d96d1aa01cb",
+      "tools/graphrag/evaluation/check_openwebui.py": "863801048231d474ba82930cf6bf6853b4410f8295b3f97fd3d5934852431fbe",
+      "tools/graphrag/evaluation/evaluate_graphrag.py": "22f48e9ac6dfae92d79ad648ca173547c6a37f98aa8199f7c37595f1b0b186f9",
+      "tools/graphrag/evaluation/evidence_contract.py": "78e24264c176e163e1782fef1634b6c45c7f31bd4692726d02a30ab45a06648e",
+      "tools/graphrag/evaluation/freeze_graphrag.py": "d8764d1aeb72208de55604785acff880002da1f818d8a8ad9477cf39e1f29efb",
+      "tools/graphrag/evaluation/graphrag_evaluator.py": "7f99bf0f97974058650b619ea49a8ee32284cb4b324f1ffab9baf54f5c4495f4",
+      "tools/graphrag/evaluation/reproducibility.py": "0e2dab6ffb17691f05f9d81dbec12ffc0f2e4e0b1a2114a75b66a5165cf7b4d2",
+      "tools/graphrag/evaluation/source_oracle.py": "aed0309435fd21bc274fde7905a0720f68f359b3c30bb0b912e3371ebf0d8072",
+      "tools/graphrag/evaluation/source_validation.py": "9671b8978623888dbed8b8b2d1b6b01c1ce73cb0a512b03919e48873990f15ed",
+      "tools/graphrag/graph_db.py": "c4b1a2edb765856b807398e4c243a8d17753529a5d2dfdb51791e28bc8de17f7",
+      "tools/graphrag/graph_search.py": "c41cbfbac0d5daa1d0b51f5f4b798a15688bbd2e0498d2cbe93c6584a99f6b7d",
+      "tools/graphrag/graphrag.py": "dbc75403aa038e4021b5a70d1f17c62784b478252a94bac8db682bf38dcb39b3",
+      "tools/graphrag/neo4j_smoke.py": "fb5244c72c29ef2acf96d1e7f54335ff168eed4119db1692025eb1b99194b7bb",
+      "tools/graphrag/publication.py": "fba1fa0d0bed325fd87f8d3593baae9271de1766b772d7214c7f63fd3f1c37b6",
+      "tools/graphrag/tests/__init__.py": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+      "tools/graphrag/tests/test_evaluation_pipeline_v2.py": "63d1b3e80ae3c7b563374fd7ab4aa5ffad3c614eafbd15142c577abd7f756bf9",
+      "tools/graphrag/tests/test_evaluation_v5.py": "9387e9d1418ebab87530986b0e29dac31cf33d762777a29a31cff24cc72e552a",
+      "tools/graphrag/tests/test_evaluation_wide.py": "1b28bb74fd6b9457c869186b88606f1f7863d573aa3ab5226162e878f236d99e",
+      "tools/graphrag/tests/test_evaluator_connected_v2.py": "29d7c9838ed1ff37ea3276df6feb1b6142158c43a0525712f490b648181882de",
+      "tools/graphrag/tests/test_graph_publication.py": "96e53c1ae28e36c7a4290e1e60c8b398bf095ceaf883e819e1a477e2672b11c0",
+      "tools/graphrag/tests/test_graph_rebuild_atomic.py": "fe37d4fff0b1f0fb1d565f0a96152b9b4e336837bc78b5977cf3b4879d7fbaf3",
+      "tools/graphrag/tests/test_graph_search.py": "eb58661aef9dafe6bf7188fcdc2d79cef6652d57c64d1b81ae81a52e8e0204d7",
+      "tools/graphrag/tests/test_graph_search_entity_v2.py": "0ef1a8ebd513c5a4507eb38fcf5ae6137c69e9901b5a2975ba513be5902ac693",
+      "tools/graphrag/tests/test_graph_search_neo4j.py": "fec2bf72de8f752f17c9164fee5754bdb2e153828855e47efbf0000dbe82cf3b",
+      "tools/graphrag/tests/test_graph_search_regressions.py": "9db7779aeb8ed9e8c0f92c1657011a81080aff3c928649e811358e3009a18c48",
+      "tools/graphrag/tests/test_graphrag.py": "9ee8629d6b528dd78c2ddfc25cee95e17613f45b454fc0a38d4c591862637a9f",
+      "tools/graphrag/tests/test_graphrag_evaluator.py": "1c6d08aeec9ae56ac4123239a5aed12a0081d66cc3f34b2047d95764579d7298",
+      "tools/graphrag/tests/test_graphrag_freeze.py": "c4b760baf4c7ebf13b16bd4cd3c51874f5512eef3b4c61470dc4d428b0303d9f",
+      "tools/graphrag/tests/test_graphrag_research_v2.py": "711008a91ab8a58ae0854e7f90329bcd42e5189fde2e2001589e97b0c22c5198",
+      "tools/graphrag/tests/test_graphrag_root_title.py": "085651fe6e6081521a5dde9554abfee93456250108cf15978df25378d886a46c",
+      "tools/graphrag/tests/test_source_oracle_v2.py": "0afdbe235e185ffe9f38aa667f90f5a6c20504fe29e9ac572642b39a8eb7faca",
+      "tools/graphrag/tests/test_title_labels.py": "2983ead68ad6e2e346d24af0fc96dc6771c163e9c2a351af426a1e0383998e6f",
+      "requirements.txt": "26a6d097727eba03c1b6c71cd02b06373ef01ce03e567dd2bc4357ae330e5a1a",
+      "tools/publication.py": "f99d445073390a733afc8d849b9fcd4b4b0890bb54d136db67d4015b78436e1a",
+      "tools/retrieval_contract.py": "b8fa272c3babf924a3ce8b937fe7ae437cd959ba9f7667193ac67c18d2c5b58d",
+      "tools/index_state.py": "c33920a63d04009f11ef23ca9e9eee9bf86453b7d0586d027bf3b39156afea0d",
+      "tools/rag/common.py": "8497c46d5c6c27a7c76df1568360ad5e9d00d5e5a7ef186e758960cc79d81d1c",
+      "tools/sync_indexes.py": "5b909b4394c24e5ce22848594b16d9aea45fb06418f5cd92bb6aa6381a0ccf10",
+      "tools/rag/rag_utils.py": "4f898cf0facc2ae9f2ea28a36f48d4ed37bd7fc716e097b319bcf4d36a4a4f03",
+      "tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
+      "tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
+      "tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
+      "tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002",
+      "tools/rag/evaluation/rag_metrics.py": "3e95922df90dd1b757a550040a2d9bd4f857e17d179f60170d7cc881a08f94b6",
+      "tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1"
+    },
+    "dataset_sha256": "0f78e9749de369d752b6e1079f174cf3f6b390f4718a672119dfd43fa5d9dfac",
+    "inputs": {
+      "data/documents.json": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+      "data/chunks.json": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9"
+    },
+    "source_corpus": {
+      "sha256": "440da12a7a991824e923f70cada7c616c3fb3b0f12767ec7d75039a6a3231b10",
+      "document_count": 114,
+      "documents": {
+        "pages/interns/bogdan_paul_chis/README.md": "3b4b946eb3c10e9f9a4092e6cffa01c5080e7579aded2f6ed3120c41ed62b647",
+        "pages/interns/cesar_gutierrez/README.md": "b0be60576eed28d4b1920e9acf198684519373dfc13171567247c853b3ed699e",
+        "pages/interns/oliver_pejic/README.md": "0470f3c8465be8f02204b08e4c75ef47f68c7e4ad1038faa5641cdf1811f33d5",
+        "pages/interns/sevval_bulburu/README.md": "ff2a9d66253802f1af0fdfd74832e40fae41b6df23c5e61dc266f6cd0cc16c4d",
+        "pages/interns/yussef_ressaissi/README.md": "f997a65699da0db3222ef7a3c37652d05dfc3419ad4ec74683bafc2c897bb987",
+        "pages/students/2016/darius_lindvai/README.md": "fcaa47bb9708a6439eec5c4b7f7b4ed6c7359e423e1d940e1aed9f44957e6d62",
+        "pages/students/2016/darius_lindvai/dp2021/README.md": "be3d0fea2a6d592b225eecabd9b19445c2fce62b8289c07ab6af93835db12970",
+        "pages/students/2016/darius_lindvai/timovy_projekt/resers/README.md": "352162de1b1e192228fa5d26506c028ac47e567e90df53b1bbaa6ad736ec4766",
+        "pages/students/2016/darius_lindvai/timovy_projekt/tutorial/README.md": "b4a81ff457624b089d618e4479afb1e3d748f6724f72104ef5518e535ae02fe5",
+        "pages/students/2016/dominik_nagy/README.md": "1c4037a158d84c45fa705a0c38e47142a9d46e226f0c138bc2d56d7c38f6d74d",
+        "pages/students/2016/dominik_nagy/timovy_projekt/README.md": "b81a4dab7316f7b846dcf36515f6aa9d44035c86f46f0e07b7a43308bcd31458",
+        "pages/students/2016/ediz_morochovic/README.md": "045e5eb9a3df2b84140808fbca0c3163266ad0efa144bc42d9d5c597881c6ac6",
+        "pages/students/2016/jakub_maruniak/README.md": "6734c41a95b9d0c1e3d226b393ad109ee4a7b82ebfffa73df2d7e0a743fefa14",
+        "pages/students/2016/jakub_maruniak/dp2021/README.md": "9cd884aec70705cdb2aaa0f514b513f16de263c37b749ced7f92916cd2f91c39",
+        "pages/students/2016/jakub_maruniak/dp2021/annotation/README.md": "debae0c50aea1accef0840912b9b1f37b2138f856002bb50e91d7ea2dbbe7963",
+        "pages/students/2016/jakub_maruniak/dp2021/train_experiments/README.md": "01dff38b37043414fa1db204394ddbe7aade00e14d02d495572b65c85d09b194",
+        "pages/students/2016/jakub_maruniak/timovy_projekt/README.md": "04114b1f3d84664aad5bdba5a0b614c5a6f23392be3c31fd9123d63db1be0a35",
+        "pages/students/2016/jan_holp/README.md": "80d6a8492766e5bfa61673a372e5f059d30c8e335de39842866c51485960e4d3",
+        "pages/students/2016/jan_holp/dp2021/README.md": "a62acb74009376766c7e2f9856ee01d21f84593eb9b3ce85cf2e6032aa412476",
+        "pages/students/2016/jan_holp/dp2021/zdrojove_subory/README.md": "e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855",
+        "pages/students/2016/jan_holp/timovy_projekt/README.md": "79eebdfd3af7e8d5c279077e285a580731b1cf8cc9b4454f6596796e9915f70c",
+        "pages/students/2016/lukas_pokryvka/README.md": "f279456c5c277ff1ee0dd89a3464ed9f9f67b44bc185a5da9811ef5cfb675002",
+        "pages/students/2016/lukas_pokryvka/dp2021/README.md": "39b61376ce108e4a930f926ab2d6a997207457121bccfcf617129a3b210a15cf",
+        "pages/students/2016/lukas_pokryvka/timovy_projekt/README.md": "4de4b6c8cfed17b6c60090f82df1d55a66be70a79d3013850689b2baebdd801b",
+        "pages/students/2016/mark_feher/README.md": "bec37c0319fe0869a0377d771db23d090dbda93c27328f5697cdaeb3bbccfe9b",
+        "pages/students/2016/maros_harahus/README.md": "0da150e91918632c9e8cd1443f04da403595a372517e254928ce9f76c9d5033d",
+        "pages/students/2016/maros_harahus/timovy_projekt/README.md": "75af19f7c68f4b646fa1dbae75f3f3dfc52633a9f9da1400d21808bb20be0d83",
+        "pages/students/2016/patrik_pavlisin/README.md": "dc1c648d7cd9b0f0e021c57f9b98189e37c100871439e7e304390715913f2763",
+        "pages/students/2016/patrik_pavlisin/dp21/README.md": "ac1eb0fd0360ad34aa21b7c8da5bb010a6edf10be5721f28dedc3986cd1d8c54",
+        "pages/students/2016/patrik_pavlisin/dp22/README.md": "f2a8c619a812b99abab618ae388eb672407950e1f33fb8d7da7a528c4de81d50",
+        "pages/students/2016/patrik_pavlisin/tp20/README.md": "ae9c771f27e7e093bc29c6888c8d0b27ed61fb41519fd0d7e7f4f6cbb4626edc",
+        "pages/students/2016/tomas_kucharik/README.md": "6a1a215433eb1857660177b4cb2ecec03c6a86743d3a04cd6d5c9d9682c9e63b",
+        "pages/students/2017/dmytro_ushatenko/README.md": "65f4ac177d492162b08ca5da3d9eb048bd088a3c117a28a662113b7e4bc3a4ec",
+        "pages/students/2017/martin_jancura/README.md": "6a0a18d5095dcfb12abfdd72c9cfd8025d714fb5e29a7858cd2b285e4cb986b9",
+        "pages/students/2017/stanislav_matsunych/README.md": "90ffa11eff4a4360b0252dbc53a116c3a5777b3d5dd86ce5c89c02bdd2051c73",
+        "pages/students/2017/than_trung_thanh/README.md": "2150519b8580e7337520db4de64d2d1c82d2634380537d019f41c506dbf2979d",
+        "pages/students/2018/NaN1name_buday/README.md": "23c8c58090bdc954347e36f141bea91f43c8d5b1d777e346864a2ebe84c0bcce",
+        "pages/students/2018/adrian_remias/README.md": "e5bfe1d5d4e626beaa636a02d34ec58cf2cff9b0e57833905af35b80bf1465a2",
+        "pages/students/2018/david_ilas/README.md": "96c45ff9c5a821ddcb1a2fc86d487bd8f4729d1b76887914b0f5fbf99af45af0",
+        "pages/students/2018/david_omasta/README.md": "4164051870e1e1cf8949ed18a74196e9ea789dfbe4fee24bdeb736ca42848327",
+        "pages/students/2018/martin_wencel/README.md": "0e6a7a981689f22304f8baa5de97d4315c0695c94c9277de7851943eb012253b",
+        "pages/students/2018/oleh_bilykh/README.md": "4072c62dd8ee753a0c8fa43fdfbbdf566e394c2df13e6a79d08a302e332710b1",
+        "pages/students/2018/ondrej_megela/README.md": "969ff2e1a436b8f957e7168cf5eb678196b99aeda17701c502b016441f644065",
+        "pages/students/2018/samuel_sirotnik/README.md": "cc230ede67b6e8632511c92a5ed6d1e9deb42e70819280222bdfd369ee77fc70",
+        "pages/students/2019/alina_vitko/README.md": "9b071fe2aff470c275ae027be236fe836f741f774cabb9cbbb55304a8d5102cd",
+        "pages/students/2019/artem_yatsenko/README.md": "328b101e46c51c959e792cf7ca226942f81d2ab8fe9648ad6d0ecd07016429e2",
+        "pages/students/2019/david_stromp/README.md": "e46ca8691d5ec894128f6f6f2ed33e2c7bd141f80719bb0449b20f10f30cca09",
+        "pages/students/2019/dmytro_mural/README.md": "e881a47d847c3ef8d680844113fdbfbb87553b28982bf9eef4936f6d30923e93",
+        "pages/students/2019/filip_tomas/README.md": "75a1fe7b921e4bff0fdf9e2a957b4adad4b4a57aa7ebdffbb73dd82c02123f94",
+        "pages/students/2019/kristian_sopkovic/README.md": "3286960fb30f584df35455f9b6dfae81ca4827a1d711d2bcd6e48c3f150c9279",
+        "pages/students/2019/michal_stromko/README.md": "3ae1dd9b6638a6a249a94be3f3846286555666aefa225bb3a2c7288b35900ce7",
+        "pages/students/2019/michal_stromko/vp2021/README.md": "feb2e093fdd6461239e79b1e1f96e65a1c8cd96e8a2b70535ae556a9c198046a",
+        "pages/students/2019/michal_stromko/vp2021/tutorials/README.md": "40e3fa448c7329ef863e705435e23fe6b4788de440710f066fe0eed92d132bad",
+        "pages/students/2019/michal_stromko/vp2021/tutorials/img/README.md": "e8d3ac160b17dfdee2c3e996bf6a03af7a6f16f9140bf5cf7ee343a82dd1865f",
+        "pages/students/2019/patrik_pokrivcak/README.md": "3e3957ad0f7aca59d6eb64c38a3f5e21d8fd953200825d0d0e5cacabfbc35a7d",
+        "pages/students/2019/samuel_horani/README.md": "86e6959ccc612b069687b57996fa3f61660075d6a0d65bfec64f423d39a0a999",
+        "pages/students/2020/david_kostilnik/README.md": "2927524b75da59a64120fc364a0fc374986fa4d364d54e52c36800762c203264",
+        "pages/students/2020/jakub_kristian_lukas/README.md": "f139d9b1ebf5a3f45612f3df5703ecdd935b90701763c27c7c529044c0b0455a",
+        "pages/students/2020/kamil_tomcufcik/README.md": "9d865adf538bc6d7992ee4f4882713447da5c936fca17c43ca0f576072537048",
+        "pages/students/2020/matej_kobylan/README.md": "3d7b289e17b8f25678a07e9ee0e6741688485a0ccb1e50940a6a6acc6312a547",
+        "pages/students/2020/matus_suchanic/README.md": "770ca49538cca22df35f8e7ba736ee97483e4fe5788c221496553592b4e78dc6",
+        "pages/students/2020/pavol_hudak/README.md": "33d06fccc6451e5f19b27b0cad001a34e81635b4ea294a01dec7771c823abc40",
+        "pages/students/2020/vladyslav_krupko/README.md": "4c28e6d38651bb2cafcbf4a18e35210cd369aef2caff6c5b0aa52fac38f0fe7d",
+        "pages/students/2021/artur_hyrenko/README.md": "0568261f4ade4de5472edcfeb1ffed9abae18dc67e4050f4ce2ad088839a2bb3",
+        "pages/students/2021/eduard_matovka/README.md": "de2a7ae38ec87110e347956b37d0605a298ed2847f972961cc9b977da8221d43",
+        "pages/students/2021/manohar_gowdru_shridharu/README.md": "c00ceb620ed7087ead5d9448d476bbb98d273d1d9e7bf1bc4355c367b26c7a6b",
+        "pages/students/2021/martin_sarissky/README.md": "8e97b45a253745faa55f63c844e0b5dfe680aeaf6eb01fb8f9d06ac6e8ee8bcf",
+        "pages/students/2021/matej_novotny/README.md": "b083a7cadde532e3428dc9a57bc61456fcc5905bda91e011b96033d42bf5d9cb",
+        "pages/students/2021/matej_scislak/README.md": "b9a32cce8605e044a4f2828ff2d91a2995a62ef3a841e62f0a225c97221f0cc8",
+        "pages/students/2021/nikita_bodnar/README.md": "2016c0dd923c0f8e561d5cea42ac81da3610fdcc54a4527f56d078e34e53ce90",
+        "pages/students/2021/oleksandr_hryshchenko/README.md": "4b073992b715b6c75b79fd2c8f850a69f940966e3a61a82c625c32c945a93036",
+        "pages/students/2021/vladimir_ferko/README.md": "68e7d311896eda2171e5eec1ec3a8da90d326ed91acaace8058f68664a50d271",
+        "pages/students/2022/andrii_pervashov/README.md": "319dcaa1cc2e29dd7e3a710bab26cb3ac6944ab71b0b97e1de076b1199122334",
+        "pages/students/2022/daniil_huzenko/README.md": "390782fbff6220a9462060ce1a62959f2a2b88f42d06e2c51d1f406b75d73f2c",
+        "pages/students/2022/jakub_schwarc/README.md": "f541a8852c2b0f6f9d7182e01a34c4e3beee9bfa55427a06b331638b11eb8be6",
+        "pages/students/2022/jan_malinovsky/README.md": "8af021bfb131fa005a7aec3e2f2f12fc17fa5dc14e2675311adbbfc445a4bcab",
+        "pages/students/2022/jan_ptak/README.md": "131cf053764fd5423541ebe672c3b4f5da875cfb8e4b821f1335ddb967282282",
+        "pages/students/2022/oleh_poiasnik/README.md": "c2f7dc3f61f8b4a5b43ded1e3ae4c7d3e7f874439ee6b17dd9c8d45dabc25bc8",
+        "pages/students/2022/samuel_vasil/README.md": "67c807049147aacd04f3050f4ff0233a8b3f0a239ab173b430e13dfe113ad0a5",
+        "pages/students/2022/serhii_yemets/README.md": "ffbeac893947acafe855d576322a8826f9d4f87ac344cb31f8ceeefa244b9133",
+        "pages/students/2022/tetiana_mohorian/README.md": "c9efdc632bf1ca89a0bfde0dc00be7aaac8a85da3d92e2ce8506236c73e93cfb",
+        "pages/students/2022/valerii_kutsenko/README.md": "dc764c609ad3c1df59cfea997d4db82a591b1b1e0764453e319d841115507e4f",
+        "pages/students/2022/vladyslav_yanchenko/README.md": "4d1ce216e3b526ee437ddbc658e51d7bb83530087e43628c096a569a3854a47f",
+        "pages/students/2022/yevhenii_leonov/README.md": "b90e41db0cec2451027020a9102d32a24d777e8a38b604ccca87b36cd6b3a292",
+        "pages/students/2023/denis_le_thanh/README.md": "2cb83f942d86a02712b0c58778e46dc2211346d1a6eba606471dfa832245db61",
+        "pages/students/2023/maksym_zatirka/README.md": "7a7d2cbe276c352dacfbc6ba6de5f15303d30ecf34845ff8da8d085f3d67c3dd",
+        "pages/students/2023/oleksandr_dorybohov/README.md": "c24d6f941de0dafb4459cea8259cf147821cafc86ba65be5b528f419f734bc18",
+        "pages/students/2023/ronald_zalacko/README.md": "724b0f8e6e641236dba9bddbd77f4546e09514076c7dfb151807aaf4b882c96f",
+        "pages/students/2023/rostyslav_rodzhuk/README.md": "de356dd7427c8a0716c165a0f88ae47b691a86ee7a388f60e2d2e9306298415c",
+        "pages/students/2023/samuel_dzurina/README.md": "4818191d044b91219c796a35010c95da5cf2093e050d494f30b219b91e13fa0e",
+        "pages/students/2023/simona_bobrovcanova/README.md": "e473a1911dae10ab40e2cf6696a8be616a58c22aaf09af5953fbac475791a35e",
+        "pages/students/2023/yevhenii_medushivskyi/README.md": "5710ebf91c893a12fe6d19f5372311938d1e7c7bbdf203f2dc8ed2fe7c4f5a24",
+        "pages/topics/akopisat/README.md": "c7245e1fbfa4160618502c0baf8cc6bc47cd9a82b8a85bc1c645314829a7b76a",
+        "pages/topics/bert/README.md": "35e507f7ec2ad1d3526e2a15f94ea085ee25f680a02daf6df62fd429bd9c25c9",
+        "pages/topics/chatbot/README.md": "6f2f83ffb80109be47f9750cca9d3036ec3645abef9f2b590bd6825780c7c48a",
+        "pages/topics/faq/README.md": "15cfbd9c2a1649c80a7a58922e2b66dadab5c86a5ef33ef16c82bebd4d2dbf06",
+        "pages/topics/hatespeech/README.md": "364df94e70dd740c63544e2586ba7b957780f714fc0c941d1b04b02233e6cbc5",
+        "pages/topics/legal/README.md": "003bb9bfe1e0ca21652091c2d11496312eb04e0689fe459e14c2d5a738e8eeea",
+        "pages/topics/morphology/README.md": "e242cc5ff1a4cc98e227566c0ead999bbc95937b7c20be62cb2afb773d0d35c0",
+        "pages/topics/named-entity/README.md": "66b2e25f70211d432d596e377d552b3b8f0c525b14c31e3bc9c956c3aeadca84",
+        "pages/topics/named-entity/navod/README.md": "1ff87079a98411de0414754c60241e08d2214ddcd63ddf974e4340cce2ce5b02",
+        "pages/topics/nlp/README.md": "6f90abf9dc3f0ff33475283c13b4a08561abf631425ec2daacb076155a9e02c2",
+        "pages/topics/otvorene/README.md": "cd5544925553fe3840c0ea6f8ef60ab756e26ab6757bf810025c948d8d4f98fc",
+        "pages/topics/prodigy/README.md": "5e4f691f2754c2ec4afbaab2f41e6db77feb5b322dd9344465eb40c35f116499",
+        "pages/topics/python/README.md": "2ef073adba672142c39e9a59fa2e00a98ab52bda54233d46fbe6940ee23f9fcc",
+        "pages/topics/question/README.md": "408078a2a67cea2699e12359e85ab641136231e6104c2afdf874a705879525df",
+        "pages/topics/question/navod/README.md": "061e737c69184b2bd244316ca7ba841f7186fe01ee80ee993742a733dcad3ecd",
+        "pages/topics/question/nezodpovedatelne/README.md": "b6703d9690f4599b973efb27595ce157773ae0fc0d55cf2c0a09d9f50e8e4c92",
+        "pages/topics/question/validacie/README.md": "3216a7127abdfc6da50ed116f44ff2d6785e7da4af70d23aceda774e745ad562",
+        "pages/topics/resources/README.md": "09405397387f00cbcc5680f7d655f2d861e4d92e5df792a62d9dfd04f3534f48",
+        "pages/topics/seq2seq/README.md": "969d15c03a63a3f8d5edc9017fe5a0fb7311fdf4de4a3473a08266241f92c1e2",
+        "pages/topics/spacy/README.md": "f504e5db77cb8694f08a57fb8636e522d59a0e669216f4bc8273bc3c5f375f87",
+        "pages/topics/steel/README.md": "b4068cef2ed197711467ac47630d5b2064fdc8828f898f7af99664e94d833330",
+        "pages/topics/translation/README.md": "a7d4289e72199b307dc5ad8fc69cda39c1a33c27a6c7447af70aa9c3cc6b8635"
+      }
+    },
+    "retrieval_policy": {
+      "published_only": true,
+      "publication_policy": "explicit-false-hidden-v1",
+      "academic_preparation_is_work": false
+    }
+  },
+  "graph": {
+    "sha256": "566817ce7ddebb2638585192c148c41421071bdf719674fc6ddddb1c99a0e36f",
+    "nodes": 446,
+    "relationships": 877
+  },
+  "parameters": {
+    "limit": 40,
+    "max_entities": 8
+  },
+  "dev_summary": {
+    "questions": 1046,
+    "source_checked_count": 1046,
+    "all_facts_supported_count": 1046,
+    "strict_count": 1046,
+    "strict_accuracy": 1.0,
+    "error_count": 0,
+    "value_score": 1.0,
+    "entity_score": 1.0,
+    "relation_score": 1.0,
+    "property_score": 1.0,
+    "relation_choice_score": 1.0,
+    "document_recall": 1.0,
+    "source_recall": 1.0,
+    "binding_score": 1.0,
+    "path_score": 1.0,
+    "answerability_accuracy": 1.0,
+    "latency_mean_seconds": 0.081848,
+    "latency_median_seconds": 0.071172,
+    "latency_p95_seconds": 0.113998
+  },
+  "source_audit_summary": {
+    "dev": {
+      "questions": 1046,
+      "source_consistent": 1046,
+      "issues": {}
+    },
+    "test": {
+      "questions": 80,
+      "source_consistent": 80,
+      "issues": {}
+    }
+  },
+  "policy": "Report all questions including negatives. No exclusions or holdout tuning.",
+  "dev_results_sha256": "d1a5d7b6d5aff3c0002f36a60cebc07627a79ec6d57de119dd587e8ab70143b0",
+  "audit_sha256": "7825d4f3e4e535eb5f49e975dfef363b9b813f1de5c97e2387e5c2f3e578e6a5"
+}

tools/graphrag/evaluation/reproducibility.py

@@ -39,6 +39,7 @@ def content_fingerprints(questions: Path, root: Path = ROOT) -> dict[str, Any]:
         'tools/index_state.py', 'tools/rag/common.py', 'tools/sync_indexes.py',
         'tools/rag/rag_utils.py', 'tools/answer_evaluation/context_policy.py',
         'tools/answer_evaluation/checked_generation.py', 'tools/answer_evaluation/answer_format.py',
+        'tools/answer_evaluation/render_answer.py',
         'tools/rag/evaluation/rag_metrics.py', 'tools/rag/evaluation/rag_runner.py'])
     return {
         'implementation': {str(path.relative_to(root)): sha256_file(path) for path in files},

tools/graphrag/evaluation/results/answers_v7_dev_2026-10-26/answers_dev.csv

@@ -0,0 +1,47 @@
+id,category,strict,error,latency_seconds,attempts,first_pass_strict
+wide-001,shared_keyword_pair,True,,7.008382885003812,1,True
+wide-002,shared_keyword_pair,True,,5.500854778001667,1,True
+wide-041,same_academic_year_pair,True,,11.343971206995775,1,True
+wide-042,same_academic_year_pair,True,,6.9671594820101745,1,True
+wide-080,same_academic_year_pair,True,,9.73526496400882,1,True
+wide-081,same_start_year_pair,True,,6.6912147020048,1,True
+wide-082,same_start_year_pair,True,,9.779698412006837,1,True
+wide-104,same_start_year_pair,True,,7.042951979994541,1,True
+wide-113,person_title_year,True,,8.861812982009724,1,True
+wide-114,person_title_year,True,,5.960790655997698,1,True
+wide-116,person_title_year,True,,5.481083346996456,1,True
+wide-119,person_title_year,True,,5.442881409995607,1,True
+wide-129,person_title_year,True,,5.64509399299277,1,True
+wide-137,person_keyword_title,True,,7.987423918006243,1,True
+wide-138,person_keyword_title,True,,6.728769060995546,1,True
+wide-139,person_keyword_title,True,,8.942678854000405,1,True
+wide-141,person_keyword_title,True,,7.0407465109892655,1,True
+wide-142,person_keyword_title,True,,7.21462029000395,1,True
+wide-143,person_keyword_title,True,,8.128857211995637,1,True
+wide-145,person_keyword_title,True,,3.167452590991161,1,True
+wide-155,person_keyword_title,True,,6.265883696993114,1,True
+wide-159,person_keyword_title,True,,9.182629139002529,1,True
+wide-161,person_keyword_title,True,,7.464124964011717,1,True
+wide-163,person_keyword_title,True,,10.472138197001186,1,True
+wide-165,person_keyword_title,True,,3.6118825629964704,1,True
+wide-169,document_author_category_multi_hop,True,,6.421270335005829,1,True
+wide-170,document_author_category_multi_hop,True,,7.26794687200163,1,True
+wide-178,document_author_category_multi_hop,True,,5.971922497003106,1,True
+wide-181,document_author_category_multi_hop,True,,6.660555914000724,1,True
+wide-193,person_multiple_categories,True,,5.167634055993403,1,True
+wide-194,person_multiple_categories,True,,5.633185668993974,1,True
+wide-209,person_multiple_categories,True,,4.679902340998524,1,True
+wide-217,candidate_set_aggregation,True,,7.1928625760047,1,True
+wide-218,candidate_set_aggregation,True,,7.220508996004355,1,True
+wide-232,candidate_set_aggregation,True,,9.97075244000007,1,True
+wide-236,candidate_set_aggregation,True,,8.210259957006201,1,True
+wide-239,candidate_set_aggregation,True,,11.907442174997414,1,True
+wide-249,conjunctive_disambiguation,True,,5.638401258998783,1,True
+wide-250,conjunctive_disambiguation,True,,9.00651249299699,1,True
+wide-283,conjunctive_disambiguation,True,,14.987600032996852,2,False
+wide-289,negative_lookup,True,,5.1113616619986715,1,True
+wide-290,negative_lookup,True,,2.8115369029983412,1,True
+v7-dev-year-1999,negative_lookup,True,,2.0942367279931204,1,True
+v7-dev-year-1899,negative_lookup,True,,2.1002495850116247,1,True
+v7-dev-year-2100,negative_lookup,True,,2.0984817599965027,1,True
+v7-dev-year-9999,negative_lookup,True,,2.9951046469941502,1,True

tools/graphrag/evaluation/results/answers_v7_dev_2026-10-26/frozen_answers.json

@@ -0,0 +1,225 @@
+{
+  "method": "Known-question development selection.",
+  "graph_freeze_sha256": "9e97794200520d0f2288e452c9e186c0291a54c63d2c1382f693a9753e4081cf",
+  "audit_sha256": "7825d4f3e4e535eb5f49e975dfef363b9b813f1de5c97e2387e5c2f3e578e6a5",
+  "generation_files": {
+    "tools/answer_evaluation/evaluate_graph_answers.py": "d81504e8fe51598c6bd1b0d482f61b9b6f51f87ba3f38b03ae1bc584675a3b76",
+    "tools/answer_evaluation/graph_answer_prompt.txt": "fc9f2f8d52e70a8d67b9f4814b3f15e6be64bf88afe53ce4db4f037bde00e064",
+    "tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
+    "tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
+    "tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
+    "tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002",
+    "tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1"
+  },
+  "model": "model120-fast",
+  "temperature": 0,
+  "workers": 3,
+  "dev_ids": [
+    "v7-dev-year-1899",
+    "v7-dev-year-1999",
+    "v7-dev-year-2100",
+    "v7-dev-year-9999",
+    "wide-001",
+    "wide-002",
+    "wide-041",
+    "wide-042",
+    "wide-080",
+    "wide-081",
+    "wide-082",
+    "wide-104",
+    "wide-113",
+    "wide-114",
+    "wide-116",
+    "wide-119",
+    "wide-129",
+    "wide-137",
+    "wide-138",
+    "wide-139",
+    "wide-141",
+    "wide-142",
+    "wide-143",
+    "wide-145",
+    "wide-155",
+    "wide-159",
+    "wide-161",
+    "wide-163",
+    "wide-165",
+    "wide-169",
+    "wide-170",
+    "wide-178",
+    "wide-181",
+    "wide-193",
+    "wide-194",
+    "wide-209",
+    "wide-217",
+    "wide-218",
+    "wide-232",
+    "wide-236",
+    "wide-239",
+    "wide-249",
+    "wide-250",
+    "wide-283",
+    "wide-289",
+    "wide-290"
+  ],
+  "test_ids": [
+    "fresh7-001",
+    "fresh7-002",
+    "fresh7-003",
+    "fresh7-004",
+    "fresh7-005",
+    "fresh7-006",
+    "fresh7-007",
+    "fresh7-008",
+    "fresh7-009",
+    "fresh7-010",
+    "fresh7-011",
+    "fresh7-012",
+    "fresh7-013",
+    "fresh7-014",
+    "fresh7-015",
+    "fresh7-016",
+    "fresh7-017",
+    "fresh7-018",
+    "fresh7-019",
+    "fresh7-020",
+    "fresh7-021",
+    "fresh7-022",
+    "fresh7-023",
+    "fresh7-024",
+    "fresh7-025",
+    "fresh7-026",
+    "fresh7-027",
+    "fresh7-028",
+    "fresh7-029",
+    "fresh7-030",
+    "fresh7-031",
+    "fresh7-032",
+    "fresh7-033",
+    "fresh7-034",
+    "fresh7-035",
+    "fresh7-036",
+    "fresh7-037",
+    "fresh7-038",
+    "fresh7-039",
+    "fresh7-040",
+    "fresh7-041",
+    "fresh7-042",
+    "fresh7-043",
+    "fresh7-044",
+    "fresh7-045",
+    "fresh7-046",
+    "fresh7-047",
+    "fresh7-048",
+    "fresh7-049",
+    "fresh7-050",
+    "fresh7-051",
+    "fresh7-052",
+    "fresh7-053",
+    "fresh7-054",
+    "fresh7-055",
+    "fresh7-056",
+    "fresh7-057",
+    "fresh7-058",
+    "fresh7-059",
+    "fresh7-060",
+    "fresh7-061",
+    "fresh7-062",
+    "fresh7-063",
+    "fresh7-064",
+    "fresh7-065",
+    "fresh7-066",
+    "fresh7-067",
+    "fresh7-068",
+    "fresh7-069",
+    "fresh7-070",
+    "fresh7-071",
+    "fresh7-072",
+    "fresh7-073",
+    "fresh7-074",
+    "fresh7-075",
+    "fresh7-076",
+    "fresh7-077",
+    "fresh7-078",
+    "fresh7-079",
+    "fresh7-080"
+  ],
+  "dev_count": 46,
+  "test_count": 80,
+  "target_accuracy": 0.92,
+  "evaluation_kind": "development",
+  "rendering": "Verified structured claims rendered deterministically; raw model outputs preserved.",
+  "selected_split": "dev",
+  "selected_ids": [
+    "wide-001",
+    "wide-002",
+    "wide-041",
+    "wide-042",
+    "wide-080",
+    "wide-081",
+    "wide-082",
+    "wide-104",
+    "wide-113",
+    "wide-114",
+    "wide-116",
+    "wide-119",
+    "wide-129",
+    "wide-137",
+    "wide-138",
+    "wide-139",
+    "wide-141",
+    "wide-142",
+    "wide-143",
+    "wide-145",
+    "wide-155",
+    "wide-159",
+    "wide-161",
+    "wide-163",
+    "wide-165",
+    "wide-169",
+    "wide-170",
+    "wide-178",
+    "wide-181",
+    "wide-193",
+    "wide-194",
+    "wide-209",
+    "wide-217",
+    "wide-218",
+    "wide-232",
+    "wide-236",
+    "wide-239",
+    "wide-249",
+    "wide-250",
+    "wide-283",
+    "wide-289",
+    "wide-290",
+    "v7-dev-year-1999",
+    "v7-dev-year-1899",
+    "v7-dev-year-2100",
+    "v7-dev-year-9999"
+  ],
+  "selected_count": 46,
+  "selected_dev_overlap": 46,
+  "corrective_request_limit": 1,
+  "format_repair_limit": 1,
+  "repair_policy": "At most one total corrective request for format, citations or retrieval-context violations. Same question/context, no gold; retain all attempts.",
+  "dev_results_sha256": null,
+  "policy": "Evaluate every declared fresh TEST case once; retain every failure; do not tune on this TEST.",
+  "selected_categories": {
+    "shared_keyword_pair": 2,
+    "same_academic_year_pair": 3,
+    "same_start_year_pair": 3,
+    "person_title_year": 5,
+    "person_keyword_title": 12,
+    "document_author_category_multi_hop": 4,
+    "person_multiple_categories": 3,
+    "candidate_set_aggregation": 5,
+    "conjunctive_disambiguation": 3,
+    "negative_lookup": 6
+  },
+  "limitations": [
+    "Known development questions; not a holdout.",
+    "Harness supplies graph context; native OpenWebUI tool choice and browser UI are outside this experiment.",
+    "Checks structured factual claims, their source support/citations and required prose values; not a general semantic judge of arbitrary prose."
+  ]
+}

tools/graphrag/evaluation/results/answers_v7_regression_2026-10-26/answers_regression.csv

@@ -0,0 +1,321 @@
+id,category,strict,error,latency_seconds,attempts,first_pass_strict
+wide-001,shared_keyword_pair,True,,5.262378749001073,1,True
+wide-002,shared_keyword_pair,True,,11.023844788011047,1,True
+wide-003,shared_keyword_pair,True,,6.837112370994873,1,True
+wide-004,shared_keyword_pair,True,,7.825073618005263,1,True
+wide-005,shared_keyword_pair,True,,10.047133992004092,1,True
+wide-006,shared_keyword_pair,True,,8.64593629699084,1,True
+wide-007,shared_keyword_pair,True,,9.661823686998105,1,True
+wide-008,shared_keyword_pair,True,,8.533432190000894,1,True
+wide-009,shared_keyword_pair,True,,8.852785049995873,1,True
+wide-010,shared_keyword_pair,True,,8.675380236003548,1,True
+wide-011,shared_keyword_pair,True,,9.178905939988908,1,True
+wide-012,shared_keyword_pair,True,,8.79627010099648,1,True
+wide-013,shared_keyword_pair,True,,9.117540171995643,1,True
+wide-014,shared_keyword_pair,True,,8.615218720005942,1,True
+wide-015,shared_keyword_pair,True,,9.395387979006045,1,True
+wide-016,shared_keyword_pair,True,,8.779454938994604,1,True
+wide-017,shared_keyword_pair,True,,9.478577678994043,1,True
+wide-018,shared_keyword_pair,True,,8.680933992000064,1,True
+wide-019,shared_keyword_pair,True,,9.240019315999234,1,True
+wide-020,shared_keyword_pair,True,,8.530438004003372,1,True
+wide-021,shared_keyword_pair,True,,9.019007034992683,1,True
+wide-022,shared_keyword_pair,True,,8.598487436000141,1,True
+wide-023,shared_keyword_pair,True,,9.186067746006302,1,True
+wide-024,shared_keyword_pair,True,,8.59094295601244,1,True
+wide-025,shared_keyword_pair,True,,8.722591666009976,1,True
+wide-026,shared_keyword_pair,True,,8.684057810998638,1,True
+wide-027,shared_keyword_pair,True,,8.90965538599994,1,True
+wide-028,shared_keyword_pair,True,,9.036431355998502,1,True
+wide-029,shared_keyword_pair,True,,16.985941207996802,2,False
+wide-030,shared_keyword_pair,True,,9.278765659008059,1,True
+wide-031,shared_keyword_pair,True,,8.859797060998972,1,True
+wide-032,shared_keyword_pair,True,,8.551712281012442,1,True
+wide-033,shared_keyword_pair,True,,8.871659873999306,1,True
+wide-034,shared_keyword_pair,True,,10.27713256199786,1,True
+wide-035,shared_keyword_pair,True,,8.731808247001027,1,True
+wide-036,shared_keyword_pair,True,,10.022244788997341,1,True
+wide-037,shared_keyword_pair,True,,7.606950429995777,1,True
+wide-038,shared_keyword_pair,True,,9.073549334003474,1,True
+wide-039,shared_keyword_pair,True,,6.649941058989498,1,True
+wide-040,shared_keyword_pair,True,,9.70009443500021,1,True
+wide-041,same_academic_year_pair,True,,4.9646135219954886,1,True
+wide-042,same_academic_year_pair,True,,8.739653348005959,1,True
+wide-043,same_academic_year_pair,True,,8.06010635799612,1,True
+wide-044,same_academic_year_pair,True,,10.40220917899569,1,True
+wide-045,same_academic_year_pair,True,,8.44839003800007,1,True
+wide-046,same_academic_year_pair,True,,9.635705161999795,1,True
+wide-047,same_academic_year_pair,True,,11.822098678007023,1,True
+wide-048,same_academic_year_pair,True,,8.148709733010037,1,True
+wide-049,same_academic_year_pair,True,,18.129530342004728,2,False
+wide-050,same_academic_year_pair,True,,8.248969623993617,1,True
+wide-051,same_academic_year_pair,True,,10.277082996006357,1,True
+wide-052,same_academic_year_pair,True,,10.281655037004384,1,True
+wide-053,same_academic_year_pair,True,,5.898193639994133,1,True
+wide-054,same_academic_year_pair,True,,7.909174476997578,1,True
+wide-055,same_academic_year_pair,True,,6.191400594994775,1,True
+wide-056,same_academic_year_pair,True,,10.229951275003259,1,True
+wide-057,same_academic_year_pair,True,,5.858902864987613,1,True
+wide-058,same_academic_year_pair,True,,9.390882232997683,1,True
+wide-059,same_academic_year_pair,True,,7.287893086002441,1,True
+wide-060,same_academic_year_pair,True,,10.045547063986305,1,True
+wide-061,same_academic_year_pair,True,,7.13498594200064,1,True
+wide-062,same_academic_year_pair,True,,8.994601780999801,1,True
+wide-063,same_academic_year_pair,True,,6.288157966002473,1,True
+wide-064,same_academic_year_pair,True,,9.025555226995493,1,True
+wide-065,same_academic_year_pair,True,,6.9650647310045315,1,True
+wide-066,same_academic_year_pair,True,,9.574780579001526,1,True
+wide-067,same_academic_year_pair,True,,8.074011475997395,1,True
+wide-068,same_academic_year_pair,True,,9.531611518992577,1,True
+wide-069,same_academic_year_pair,True,,8.634510207004496,1,True
+wide-070,same_academic_year_pair,True,,8.958400035000523,1,True
+wide-071,same_academic_year_pair,True,,9.77839753999433,1,True
+wide-072,same_academic_year_pair,True,,8.366447199005052,1,True
+wide-073,same_academic_year_pair,True,,9.756975818992942,1,True
+wide-074,same_academic_year_pair,True,,7.232408730007592,1,True
+wide-075,same_academic_year_pair,True,,9.765164623997407,1,True
+wide-076,same_academic_year_pair,True,,7.285313993008458,1,True
+wide-077,same_academic_year_pair,True,,10.151919288997306,1,True
+wide-078,same_academic_year_pair,True,,7.673378441002569,1,True
+wide-079,same_academic_year_pair,True,,10.634893157999613,1,True
+wide-080,same_academic_year_pair,True,,4.789500615996076,1,True
+wide-081,same_start_year_pair,True,,8.585048984998139,1,True
+wide-082,same_start_year_pair,True,,8.499404791000416,1,True
+wide-083,same_start_year_pair,True,,5.630402820999734,1,True
+wide-084,same_start_year_pair,True,,6.908654918996035,1,True
+wide-085,same_start_year_pair,True,,5.944816776987864,1,True
+wide-086,same_start_year_pair,True,,9.126075861990103,1,True
+wide-087,same_start_year_pair,True,,5.440862012997968,1,True
+wide-088,same_start_year_pair,True,,6.5059300099965185,1,True
+wide-089,same_start_year_pair,True,,6.697157860995503,1,True
+wide-090,same_start_year_pair,True,,3.26202636299422,1,True
+wide-091,same_start_year_pair,True,,6.4398382079962175,1,True
+wide-092,same_start_year_pair,True,,9.066495395993115,1,True
+wide-093,same_start_year_pair,True,,7.360152271998231,1,True
+wide-094,same_start_year_pair,True,,9.541269890003605,1,True
+wide-095,same_start_year_pair,True,,7.569437859012396,1,True
+wide-096,same_start_year_pair,True,,11.024270423004054,1,True
+wide-097,same_start_year_pair,True,,5.715482076993794,1,True
+wide-098,same_start_year_pair,True,,8.736371631996008,1,True
+wide-099,same_start_year_pair,True,,6.249446924994118,1,True
+wide-100,same_start_year_pair,True,,10.805812649006839,1,True
+wide-101,same_start_year_pair,True,,5.220276794992969,1,True
+wide-102,same_start_year_pair,True,,9.947560219006846,1,True
+wide-103,same_start_year_pair,True,,7.779931154000224,1,True
+wide-104,same_start_year_pair,True,,10.890818430008949,1,True
+wide-105,same_start_year_pair,True,,3.6545174839993706,1,True
+wide-106,same_start_year_pair,True,,7.7419946619920665,1,True
+wide-107,same_start_year_pair,True,,4.410995238999021,1,True
+wide-108,same_start_year_pair,True,,6.583789386000717,1,True
+wide-109,same_start_year_pair,True,,6.958590795999044,1,True
+wide-110,same_start_year_pair,True,,11.648579076994793,1,True
+wide-111,same_start_year_pair,True,,3.022186320988112,1,True
+wide-112,same_start_year_pair,True,,2.6744268070033286,1,True
+wide-113,person_title_year,True,,3.9001984620117582,1,True
+wide-114,person_title_year,True,,6.90090462200169,1,True
+wide-115,person_title_year,True,,5.2774354059947655,1,True
+wide-116,person_title_year,True,,5.405227045004722,1,True
+wide-117,person_title_year,True,,4.662688102005632,1,True
+wide-118,person_title_year,True,,4.6598335029993905,1,True
+wide-119,person_title_year,True,,7.607376700005261,1,True
+wide-120,person_title_year,True,,3.551628603003337,1,True
+wide-121,person_title_year,True,,6.179311268002493,1,True
+wide-122,person_title_year,True,,5.917100324993953,1,True
+wide-123,person_title_year,True,,6.548046129988506,1,True
+wide-124,person_title_year,True,,7.861738689010963,1,True
+wide-125,person_title_year,True,,4.563962607993744,1,True
+wide-126,person_title_year,True,,5.80828094399476,1,True
+wide-127,person_title_year,True,,5.077105101998313,1,True
+wide-128,person_title_year,True,,6.038464580007712,1,True
+wide-129,person_title_year,True,,7.096872989000985,1,True
+wide-130,person_title_year,True,,7.872525553000742,2,False
+wide-131,person_title_year,True,,7.498921799997333,1,True
+wide-132,person_title_year,True,,5.892531138000777,1,True
+wide-133,person_title_year,True,,5.6638127920014085,1,True
+wide-134,person_title_year,True,,5.172251234005671,1,True
+wide-135,person_title_year,True,,5.744371400011005,1,True
+wide-136,person_title_year,True,,4.798829792998731,1,True
+wide-137,person_keyword_title,True,,6.910317715010024,1,True
+wide-138,person_keyword_title,True,,6.913008176998119,1,True
+wide-139,person_keyword_title,True,,8.655944869999075,1,True
+wide-140,person_keyword_title,True,,7.773692659000517,1,True
+wide-141,person_keyword_title,True,,9.624433186996612,1,True
+wide-142,person_keyword_title,True,,3.844812393988832,1,True
+wide-143,person_keyword_title,True,,7.311962670995854,1,True
+wide-144,person_keyword_title,True,,8.996415396002703,1,True
+wide-145,person_keyword_title,True,,4.76172783500806,1,True
+wide-146,person_keyword_title,True,,6.710666915998445,1,True
+wide-147,person_keyword_title,True,,7.200240036996547,1,True
+wide-148,person_keyword_title,True,,8.873094891998335,1,True
+wide-149,person_keyword_title,True,,7.56319696000719,1,True
+wide-150,person_keyword_title,True,,8.975214990001405,1,True
+wide-151,person_keyword_title,True,,7.537857122006244,1,True
+wide-152,person_keyword_title,True,,8.62167695100652,1,True
+wide-153,person_keyword_title,True,,6.811467941006413,1,True
+wide-154,person_keyword_title,True,,9.136220678992686,1,True
+wide-155,person_keyword_title,True,,7.490241913008504,1,True
+wide-156,person_keyword_title,True,,9.619424235992483,1,True
+wide-157,person_keyword_title,True,,7.075699581997469,1,True
+wide-158,person_keyword_title,True,,9.402772761997767,1,True
+wide-159,person_keyword_title,True,,7.3926155700028175,1,True
+wide-160,person_keyword_title,True,,9.592219272002694,1,True
+wide-161,person_keyword_title,True,,7.692930470002466,1,True
+wide-162,person_keyword_title,True,,9.336959647000185,1,True
+wide-163,person_keyword_title,True,,7.659145907993661,1,True
+wide-164,person_keyword_title,True,,9.989820002010674,1,True
+wide-165,person_keyword_title,True,,3.6757876240008045,1,True
+wide-166,person_keyword_title,True,,6.71545403500204,1,True
+wide-167,person_keyword_title,True,,8.453232906991616,1,True
+wide-168,person_keyword_title,True,,7.789937220004504,1,True
+wide-169,document_author_category_multi_hop,True,,8.230146304005757,1,True
+wide-170,document_author_category_multi_hop,True,,6.4932332140015205,1,True
+wide-171,document_author_category_multi_hop,True,,7.096607529994799,1,True
+wide-172,document_author_category_multi_hop,True,,5.593765233003069,1,True
+wide-173,document_author_category_multi_hop,True,,6.806600865995279,1,True
+wide-174,document_author_category_multi_hop,True,,5.3536878140002955,1,True
+wide-175,document_author_category_multi_hop,True,,7.072415729999193,1,True
+wide-176,document_author_category_multi_hop,True,,6.418687598998076,1,True
+wide-177,document_author_category_multi_hop,True,,7.768641275004484,1,True
+wide-178,document_author_category_multi_hop,True,,6.695670635002898,1,True
+wide-179,document_author_category_multi_hop,True,,6.646737192000728,1,True
+wide-180,document_author_category_multi_hop,True,,5.989018792010029,1,True
+wide-181,document_author_category_multi_hop,True,,6.119498893007403,1,True
+wide-182,document_author_category_multi_hop,True,,5.9745991970121395,1,True
+wide-183,document_author_category_multi_hop,True,,6.077126152988058,1,True
+wide-184,document_author_category_multi_hop,True,,5.992435852997005,1,True
+wide-185,document_author_category_multi_hop,True,,6.213222564998432,1,True
+wide-186,document_author_category_multi_hop,True,,6.683954360996722,1,True
+wide-187,document_author_category_multi_hop,True,,6.843328906004899,1,True
+wide-188,document_author_category_multi_hop,True,,7.07258083600027,1,True
+wide-189,document_author_category_multi_hop,True,,6.259952289008652,1,True
+wide-190,document_author_category_multi_hop,True,,6.505296975999954,1,True
+wide-191,document_author_category_multi_hop,True,,6.268639836998773,1,True
+wide-192,document_author_category_multi_hop,True,,7.003398792003281,1,True
+wide-193,person_multiple_categories,True,,6.194855802998063,1,True
+wide-194,person_multiple_categories,True,,5.364437454001745,1,True
+wide-195,person_multiple_categories,True,,4.545286751992535,1,True
+wide-196,person_multiple_categories,True,,4.151337641000282,1,True
+wide-197,person_multiple_categories,True,,8.146557949992712,2,False
+wide-198,person_multiple_categories,True,,4.233387041007518,1,True
+wide-199,person_multiple_categories,True,,6.291938512003981,1,True
+wide-200,person_multiple_categories,True,,9.257844247011235,2,False
+wide-201,person_multiple_categories,True,,3.8606335170043167,1,True
+wide-202,person_multiple_categories,True,,4.706902887002798,1,True
+wide-203,person_multiple_categories,True,,4.9716377110016765,1,True
+wide-204,person_multiple_categories,True,,11.950999951994163,2,False
+wide-205,person_multiple_categories,True,,5.60192243399797,1,True
+wide-206,person_multiple_categories,True,,5.590988194002421,1,True
+wide-207,person_multiple_categories,True,,7.846379655005876,2,False
+wide-208,person_multiple_categories,True,,7.050330217010924,1,True
+wide-209,person_multiple_categories,True,,5.981790188001469,1,True
+wide-210,person_multiple_categories,True,,4.014929401004338,1,True
+wide-211,person_multiple_categories,True,,5.048257627000567,1,True
+wide-212,person_multiple_categories,True,,2.727644105994841,1,True
+wide-213,person_multiple_categories,True,,4.587378602998797,1,True
+wide-214,person_multiple_categories,True,,3.6640594680065988,1,True
+wide-215,person_multiple_categories,True,,7.165378771998803,1,True
+wide-216,person_multiple_categories,True,,6.913648023997666,2,False
+wide-217,candidate_set_aggregation,True,,8.149436000996502,1,True
+wide-218,candidate_set_aggregation,True,,7.928223324997816,1,True
+wide-219,candidate_set_aggregation,True,,8.299362689998816,1,True
+wide-220,candidate_set_aggregation,True,,9.450791957002366,1,True
+wide-221,candidate_set_aggregation,True,,8.832846570992842,1,True
+wide-222,candidate_set_aggregation,True,,9.567884083997342,1,True
+wide-223,candidate_set_aggregation,True,,8.838387846000842,1,True
+wide-224,candidate_set_aggregation,True,,9.386160062989802,1,True
+wide-225,candidate_set_aggregation,True,,8.62494325399166,1,True
+wide-226,candidate_set_aggregation,True,,9.428272984994692,1,True
+wide-227,candidate_set_aggregation,True,,8.599248733007698,1,True
+wide-228,candidate_set_aggregation,True,,9.487015202001203,1,True
+wide-229,candidate_set_aggregation,True,,8.605383807007456,1,True
+wide-230,candidate_set_aggregation,True,,9.61059859499801,1,True
+wide-231,candidate_set_aggregation,True,,8.823051813000347,1,True
+wide-232,candidate_set_aggregation,True,,9.740953706990695,1,True
+wide-233,candidate_set_aggregation,True,,8.722386483990704,1,True
+wide-234,candidate_set_aggregation,True,,9.489644886998576,1,True
+wide-235,candidate_set_aggregation,True,,8.559914067009231,1,True
+wide-236,candidate_set_aggregation,True,,9.524188774012146,1,True
+wide-237,candidate_set_aggregation,True,,8.78702475198952,1,True
+wide-238,candidate_set_aggregation,True,,9.779589921003208,1,True
+wide-239,candidate_set_aggregation,True,,8.832212022010935,1,True
+wide-240,candidate_set_aggregation,True,,9.979131497995695,1,True
+wide-241,candidate_set_aggregation,True,,8.59488222800428,1,True
+wide-242,candidate_set_aggregation,True,,9.823719104999327,1,True
+wide-243,candidate_set_aggregation,True,,8.220046620001085,1,True
+wide-244,candidate_set_aggregation,True,,10.194505384002696,1,True
+wide-245,candidate_set_aggregation,True,,8.194634853003663,1,True
+wide-246,candidate_set_aggregation,True,,10.07240725599695,1,True
+wide-247,candidate_set_aggregation,True,,7.7612789029954,1,True
+wide-248,candidate_set_aggregation,True,,11.560260020007263,1,True
+wide-249,conjunctive_disambiguation,True,,5.476591879996704,1,True
+wide-250,conjunctive_disambiguation,True,,8.843467402999522,1,True
+wide-251,conjunctive_disambiguation,True,,7.115450201003114,1,True
+wide-252,conjunctive_disambiguation,True,,8.416170042997692,1,True
+wide-253,conjunctive_disambiguation,True,,6.920364450008492,1,True
+wide-254,conjunctive_disambiguation,True,,8.249018574002548,1,True
+wide-255,conjunctive_disambiguation,True,,6.858485064003617,1,True
+wide-256,conjunctive_disambiguation,True,,8.336330271995394,1,True
+wide-257,conjunctive_disambiguation,True,,6.96214788701036,1,True
+wide-258,conjunctive_disambiguation,True,,8.381303394009592,1,True
+wide-259,conjunctive_disambiguation,True,,7.045915052003693,1,True
+wide-260,conjunctive_disambiguation,True,,8.721610349006369,1,True
+wide-261,conjunctive_disambiguation,True,,6.935596920011449,1,True
+wide-262,conjunctive_disambiguation,True,,8.776731780002592,1,True
+wide-263,conjunctive_disambiguation,True,,6.759910934997606,1,True
+wide-264,conjunctive_disambiguation,True,,9.799247552000452,1,True
+wide-265,conjunctive_disambiguation,True,,5.189480711997021,1,True
+wide-266,conjunctive_disambiguation,True,,8.85979151999345,1,True
+wide-267,conjunctive_disambiguation,True,,6.7524094420077745,1,True
+wide-268,conjunctive_disambiguation,True,,9.291388099998585,1,True
+wide-269,conjunctive_disambiguation,True,,6.566982984004426,1,True
+wide-270,conjunctive_disambiguation,True,,8.416923523996957,1,True
+wide-271,conjunctive_disambiguation,True,,6.700699591994635,1,True
+wide-272,conjunctive_disambiguation,True,,8.531122063999646,1,True
+wide-273,conjunctive_disambiguation,True,,6.868541053991066,1,True
+wide-274,conjunctive_disambiguation,True,,8.585297629004344,1,True
+wide-275,conjunctive_disambiguation,True,,6.873942530000932,1,True
+wide-276,conjunctive_disambiguation,True,,8.594346833997406,1,True
+wide-277,conjunctive_disambiguation,True,,6.816038967008353,1,True
+wide-278,conjunctive_disambiguation,True,,8.591339434002293,1,True
+wide-279,conjunctive_disambiguation,True,,6.7824973950046115,1,True
+wide-280,conjunctive_disambiguation,True,,8.697670733003179,1,True
+wide-281,conjunctive_disambiguation,True,,6.762842822994571,1,True
+wide-282,conjunctive_disambiguation,True,,8.647353213993483,1,True
+wide-283,conjunctive_disambiguation,True,,18.08595217599941,2,False
+wide-284,conjunctive_disambiguation,True,,7.201167734005139,1,True
+wide-285,conjunctive_disambiguation,False,,28.01832078601001,2,False
+wide-286,conjunctive_disambiguation,True,,5.158636806998402,1,True
+wide-287,conjunctive_disambiguation,True,,11.181534039991675,1,True
+wide-288,conjunctive_disambiguation,True,,5.375784352989285,1,True
+wide-289,negative_lookup,True,,3.1519780079979682,1,True
+wide-290,negative_lookup,True,,2.5652610169927357,1,True
+wide-291,negative_lookup,True,,2.116932502001873,1,True
+wide-292,negative_lookup,True,,2.1522221219929634,1,True
+wide-293,negative_lookup,True,,2.2185270579939242,1,True
+wide-294,negative_lookup,True,,2.3594186319969594,1,True
+wide-295,negative_lookup,True,,3.4991669790033484,1,True
+wide-296,negative_lookup,True,,3.309362318002968,1,True
+wide-297,negative_lookup,True,,3.1218993680085987,1,True
+wide-298,negative_lookup,True,,2.7141988500079606,1,True
+wide-299,negative_lookup,True,,2.4447732209955575,1,True
+wide-300,negative_lookup,True,,2.579182736008079,1,True
+wide-301,negative_lookup,True,,2.425772479007719,1,True
+wide-302,negative_lookup,True,,2.5622203609964345,1,True
+wide-303,negative_lookup,True,,2.476981298008468,1,True
+wide-304,negative_lookup,True,,2.6007943300064653,1,True
+wide-305,negative_lookup,True,,2.457786357990699,1,True
+wide-306,negative_lookup,True,,2.566486335999798,1,True
+wide-307,negative_lookup,True,,2.6169559550035046,1,True
+wide-308,negative_lookup,True,,2.7340419859974645,1,True
+wide-309,negative_lookup,True,,2.619804314992507,1,True
+wide-310,negative_lookup,True,,2.556121910994989,1,True
+wide-311,negative_lookup,True,,2.4317323060095077,1,True
+wide-312,negative_lookup,True,,2.506849172001239,1,True
+wide-313,negative_lookup,True,,2.4141654220002238,1,True
+wide-314,negative_lookup,True,,2.4877845580049325,1,True
+wide-315,negative_lookup,True,,2.4770866070030024,1,True
+wide-316,negative_lookup,True,,2.494239155013929,1,True
+wide-317,negative_lookup,True,,2.461824482001248,1,True
+wide-318,negative_lookup,True,,2.4353119149891427,1,True
+wide-319,negative_lookup,True,,2.512933713005623,1,True
+wide-320,negative_lookup,True,,1.8379592370038154,1,True

tools/graphrag/evaluation/results/answers_v7_regression_2026-10-26/failures.json

@@ -0,0 +1,260 @@
+{
+  "evaluation_kind": "historical_regression",
+  "questions": [
+    {
+      "id": "wide-285",
+      "category": "conjunctive_disambiguation",
+      "question": "Ktorá z osôb Dárius Lindvai a Dominik Nagy má kľúčové slovo pytorch a jej práca je z roku 2019?",
+      "graph_context": {
+        "status": "ok",
+        "reason": null,
+        "facts": [
+          {
+            "person": "Dárius Lindvai",
+            "year": 2019,
+            "document_path": "pages/students/2016/darius_lindvai/README.md",
+            "source_url": "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai"
+          },
+          {
+            "person": "Dárius Lindvai",
+            "keyword": "pytorch",
+            "document_path": "pages/students/2016/darius_lindvai/timovy_projekt/tutorial/README.md",
+            "source_url": "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial"
+          }
+        ],
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "pages/students/2016/darius_lindvai/README.md",
+            "source_url": "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai",
+            "title": "Dárius Lindvai"
+          },
+          {
+            "source_id": "S2",
+            "document_path": "pages/students/2016/darius_lindvai/timovy_projekt/tutorial/README.md",
+            "source_url": "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial",
+            "title": "PYTORCH - LSTM TUTORIÁL"
+          },
+          {
+            "source_id": "S3",
+            "document_path": "pages/students/2016/dominik_nagy/README.md",
+            "source_url": "https://zp.kemt.fei.tuke.sk/students/2016/dominik_nagy",
+            "title": "Dominik Nagy"
+          }
+        ],
+        "aggregation": null,
+        "complete": true,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+          "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+          "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+          "embedding_model": "intfloat/multilingual-e5-small",
+          "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+          "publication_policy": "explicit-false-hidden-v1",
+          "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+          "version": "33ea0794472c416e97c84371a2a6cdf8"
+        },
+        "answer_candidates": [
+          {
+            "person": "Dárius Lindvai",
+            "year": 2019,
+            "keyword": "pytorch",
+            "evidence_documents": [
+              "pages/students/2016/darius_lindvai/README.md",
+              "pages/students/2016/darius_lindvai/timovy_projekt/tutorial/README.md"
+            ],
+            "source_urls": [
+              "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai",
+              "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial"
+            ]
+          }
+        ]
+      },
+      "strict": false,
+      "attempts": [
+        {
+          "response": {
+            "choices": [
+              {
+                "finish_reason": "stop",
+                "index": 0,
+                "message": {
+                  "role": "assistant",
+                  "content": "{\n  \"answer\": \"\",\n  \"status\": \"ok\",\n  \"claims\": [\n    {\n      \"person\": \"Dárius Lindvai\",\n      \"year\": 2019,\n      \"keyword\": \"pytorch\",\n      \"title\": \"PYTORCH - LSTM TUTORIÁL\",\n      \"source_urls\": [\n        \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai\",\n        \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial\"\n      ]\n    }\n  ],\n  \"source_urls\": [\n    \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai\",\n    \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial\"\n  ],\n  \"count\": null\n}"
+                }
+              }
+            ],
+            "created": 1792974833,
+            "model": "Qwen3.5-122B-A10B-Q4_K_M",
+            "system_fingerprint": "b8580-7c203670f",
+            "object": "chat.completion",
+            "usage": {
+              "completion_tokens": 223,
+              "prompt_tokens": 1772,
+              "total_tokens": 1995,
+              "prompt_tokens_details": {
+                "cached_tokens": 815
+              }
+            },
+            "id": "chatcmpl-PTuYHRUKhpKd7EhYb0cDozPNHqVaQYxm",
+            "timings": {
+              "cache_n": 815,
+              "prompt_n": 957,
+              "prompt_ms": 2525.225,
+              "prompt_per_token_ms": 2.6386886102403344,
+              "prompt_per_second": 378.97613083982617,
+              "predicted_n": 223,
+              "predicted_ms": 7961.962,
+              "predicted_per_token_ms": 35.70386547085202,
+              "predicted_per_second": 28.00817185512817
+            }
+          },
+          "format_error": "ValidationError",
+          "strict": false
+        },
+        {
+          "response": {
+            "choices": [
+              {
+                "finish_reason": "stop",
+                "index": 0,
+                "message": {
+                  "role": "assistant",
+                  "content": "{\n  \"answer\": \"\",\n  \"status\": \"ok\",\n  \"claims\": [\n    {\n      \"person\": \"Dárius Lindvai\",\n      \"title\": \"PYTORCH - LSTM TUTORIÁL\",\n      \"year\": 2019,\n      \"keyword\": \"pytorch\"\n    }\n  ],\n  \"source_urls\": [\n    \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai\",\n    \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial\"\n  ],\n  \"count\": null\n}"
+                }
+              }
+            ],
+            "created": 1792974847,
+            "model": "Qwen3.5-122B-A10B-Q4_K_M",
+            "system_fingerprint": "b8580-7c203670f",
+            "object": "chat.completion",
+            "usage": {
+              "completion_tokens": 155,
+              "prompt_tokens": 2444,
+              "total_tokens": 2599,
+              "prompt_tokens_details": {
+                "cached_tokens": 1768
+              }
+            },
+            "id": "chatcmpl-EqJeUwzc7k5kwW0XpAbdx69edFgRW16H",
+            "timings": {
+              "cache_n": 1768,
+              "prompt_n": 676,
+              "prompt_ms": 946.839,
+              "prompt_per_token_ms": 1.4006494082840237,
+              "prompt_per_second": 713.9545371493991,
+              "predicted_n": 155,
+              "predicted_ms": 8280.544,
+              "predicted_per_token_ms": 53.42286451612903,
+              "predicted_per_second": 18.71857694373703
+            }
+          },
+          "answer": {
+            "answer": "",
+            "status": "ok",
+            "claims": [
+              {
+                "person": "Dárius Lindvai",
+                "title": "PYTORCH - LSTM TUTORIÁL",
+                "year": 2019,
+                "keyword": "pytorch"
+              }
+            ],
+            "source_urls": [
+              "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai",
+              "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial"
+            ]
+          },
+          "format_normalized": false,
+          "context_issues": [
+            "unsupported_academic_binding"
+          ],
+          "strict": false
+        }
+      ],
+      "accepted_for_delivery": false,
+      "response": {
+        "choices": [
+          {
+            "finish_reason": "stop",
+            "index": 0,
+            "message": {
+              "role": "assistant",
+              "content": "{\n  \"answer\": \"\",\n  \"status\": \"ok\",\n  \"claims\": [\n    {\n      \"person\": \"Dárius Lindvai\",\n      \"title\": \"PYTORCH - LSTM TUTORIÁL\",\n      \"year\": 2019,\n      \"keyword\": \"pytorch\"\n    }\n  ],\n  \"source_urls\": [\n    \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai\",\n    \"https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial\"\n  ],\n  \"count\": null\n}"
+            }
+          }
+        ],
+        "created": 1792974847,
+        "model": "Qwen3.5-122B-A10B-Q4_K_M",
+        "system_fingerprint": "b8580-7c203670f",
+        "object": "chat.completion",
+        "usage": {
+          "completion_tokens": 155,
+          "prompt_tokens": 2444,
+          "total_tokens": 2599,
+          "prompt_tokens_details": {
+            "cached_tokens": 1768
+          }
+        },
+        "id": "chatcmpl-EqJeUwzc7k5kwW0XpAbdx69edFgRW16H",
+        "timings": {
+          "cache_n": 1768,
+          "prompt_n": 676,
+          "prompt_ms": 946.839,
+          "prompt_per_token_ms": 1.4006494082840237,
+          "prompt_per_second": 713.9545371493991,
+          "predicted_n": 155,
+          "predicted_ms": 8280.544,
+          "predicted_per_token_ms": 53.42286451612903,
+          "predicted_per_second": 18.71857694373703
+        }
+      },
+      "answer": {
+        "answer": "",
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Dárius Lindvai",
+            "title": "PYTORCH - LSTM TUTORIÁL",
+            "year": 2019,
+            "keyword": "pytorch"
+          }
+        ],
+        "source_urls": [
+          "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai",
+          "https://zp.kemt.fei.tuke.sk/students/2016/darius_lindvai/timovy_projekt/tutorial"
+        ]
+      },
+      "context_rejected": true,
+      "bindings_ok": true,
+      "unsupported_claims": [
+        {
+          "person": "Dárius Lindvai",
+          "title": "PYTORCH - LSTM TUTORIÁL",
+          "year": 2019,
+          "keyword": "pytorch"
+        }
+      ],
+      "sources_ok": true,
+      "status_ok": true,
+      "count_ok": true,
+      "set_ok": true,
+      "citations_ok": true,
+      "unsupported_urls": [],
+      "context_issues": [
+        "empty_answer_text",
+        "unsupported_academic_binding"
+      ],
+      "text_values_ok": false,
+      "missing_text_values": [
+        "Dárius Lindvai",
+        "pytorch",
+        "2019"
+      ],
+      "contradictory_thesis_label": false,
+      "latency_seconds": 28.01832078601001
+    }
+  ]
+}

tools/graphrag/evaluation/results/answers_v7_regression_2026-10-26/frozen_answers.json

@@ -0,0 +1,499 @@
+{
+  "method": "Previously exposed LLM questions; regression, not independent TEST.",
+  "graph_freeze_sha256": "9e97794200520d0f2288e452c9e186c0291a54c63d2c1382f693a9753e4081cf",
+  "audit_sha256": "7825d4f3e4e535eb5f49e975dfef363b9b813f1de5c97e2387e5c2f3e578e6a5",
+  "generation_files": {
+    "tools/answer_evaluation/evaluate_graph_answers.py": "d81504e8fe51598c6bd1b0d482f61b9b6f51f87ba3f38b03ae1bc584675a3b76",
+    "tools/answer_evaluation/graph_answer_prompt.txt": "fc9f2f8d52e70a8d67b9f4814b3f15e6be64bf88afe53ce4db4f037bde00e064",
+    "tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
+    "tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
+    "tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
+    "tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002",
+    "tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1"
+  },
+  "model": "model120-fast",
+  "temperature": 0,
+  "workers": 3,
+  "dev_ids": [
+    "v7-dev-year-1899",
+    "v7-dev-year-1999",
+    "v7-dev-year-2100",
+    "v7-dev-year-9999",
+    "wide-001",
+    "wide-002",
+    "wide-041",
+    "wide-042",
+    "wide-080",
+    "wide-081",
+    "wide-082",
+    "wide-104",
+    "wide-113",
+    "wide-114",
+    "wide-116",
+    "wide-119",
+    "wide-129",
+    "wide-137",
+    "wide-138",
+    "wide-139",
+    "wide-141",
+    "wide-142",
+    "wide-143",
+    "wide-145",
+    "wide-155",
+    "wide-159",
+    "wide-161",
+    "wide-163",
+    "wide-165",
+    "wide-169",
+    "wide-170",
+    "wide-178",
+    "wide-181",
+    "wide-193",
+    "wide-194",
+    "wide-209",
+    "wide-217",
+    "wide-218",
+    "wide-232",
+    "wide-236",
+    "wide-239",
+    "wide-249",
+    "wide-250",
+    "wide-283",
+    "wide-289",
+    "wide-290"
+  ],
+  "test_ids": [
+    "fresh7-001",
+    "fresh7-002",
+    "fresh7-003",
+    "fresh7-004",
+    "fresh7-005",
+    "fresh7-006",
+    "fresh7-007",
+    "fresh7-008",
+    "fresh7-009",
+    "fresh7-010",
+    "fresh7-011",
+    "fresh7-012",
+    "fresh7-013",
+    "fresh7-014",
+    "fresh7-015",
+    "fresh7-016",
+    "fresh7-017",
+    "fresh7-018",
+    "fresh7-019",
+    "fresh7-020",
+    "fresh7-021",
+    "fresh7-022",
+    "fresh7-023",
+    "fresh7-024",
+    "fresh7-025",
+    "fresh7-026",
+    "fresh7-027",
+    "fresh7-028",
+    "fresh7-029",
+    "fresh7-030",
+    "fresh7-031",
+    "fresh7-032",
+    "fresh7-033",
+    "fresh7-034",
+    "fresh7-035",
+    "fresh7-036",
+    "fresh7-037",
+    "fresh7-038",
+    "fresh7-039",
+    "fresh7-040",
+    "fresh7-041",
+    "fresh7-042",
+    "fresh7-043",
+    "fresh7-044",
+    "fresh7-045",
+    "fresh7-046",
+    "fresh7-047",
+    "fresh7-048",
+    "fresh7-049",
+    "fresh7-050",
+    "fresh7-051",
+    "fresh7-052",
+    "fresh7-053",
+    "fresh7-054",
+    "fresh7-055",
+    "fresh7-056",
+    "fresh7-057",
+    "fresh7-058",
+    "fresh7-059",
+    "fresh7-060",
+    "fresh7-061",
+    "fresh7-062",
+    "fresh7-063",
+    "fresh7-064",
+    "fresh7-065",
+    "fresh7-066",
+    "fresh7-067",
+    "fresh7-068",
+    "fresh7-069",
+    "fresh7-070",
+    "fresh7-071",
+    "fresh7-072",
+    "fresh7-073",
+    "fresh7-074",
+    "fresh7-075",
+    "fresh7-076",
+    "fresh7-077",
+    "fresh7-078",
+    "fresh7-079",
+    "fresh7-080"
+  ],
+  "dev_count": 46,
+  "test_count": 80,
+  "target_accuracy": 0.92,
+  "evaluation_kind": "historical_regression",
+  "rendering": "Verified structured claims rendered deterministically; raw model outputs preserved.",
+  "selected_split": "regression",
+  "selected_ids": [
+    "wide-001",
+    "wide-002",
+    "wide-003",
+    "wide-004",
+    "wide-005",
+    "wide-006",
+    "wide-007",
+    "wide-008",
+    "wide-009",
+    "wide-010",
+    "wide-011",
+    "wide-012",
+    "wide-013",
+    "wide-014",
+    "wide-015",
+    "wide-016",
+    "wide-017",
+    "wide-018",
+    "wide-019",
+    "wide-020",
+    "wide-021",
+    "wide-022",
+    "wide-023",
+    "wide-024",
+    "wide-025",
+    "wide-026",
+    "wide-027",
+    "wide-028",
+    "wide-029",
+    "wide-030",
+    "wide-031",
+    "wide-032",
+    "wide-033",
+    "wide-034",
+    "wide-035",
+    "wide-036",
+    "wide-037",
+    "wide-038",
+    "wide-039",
+    "wide-040",
+    "wide-041",
+    "wide-042",
+    "wide-043",
+    "wide-044",
+    "wide-045",
+    "wide-046",
+    "wide-047",
+    "wide-048",
+    "wide-049",
+    "wide-050",
+    "wide-051",
+    "wide-052",
+    "wide-053",
+    "wide-054",
+    "wide-055",
+    "wide-056",
+    "wide-057",
+    "wide-058",
+    "wide-059",
+    "wide-060",
+    "wide-061",
+    "wide-062",
+    "wide-063",
+    "wide-064",
+    "wide-065",
+    "wide-066",
+    "wide-067",
+    "wide-068",
+    "wide-069",
+    "wide-070",
+    "wide-071",
+    "wide-072",
+    "wide-073",
+    "wide-074",
+    "wide-075",
+    "wide-076",
+    "wide-077",
+    "wide-078",
+    "wide-079",
+    "wide-080",
+    "wide-081",
+    "wide-082",
+    "wide-083",
+    "wide-084",
+    "wide-085",
+    "wide-086",
+    "wide-087",
+    "wide-088",
+    "wide-089",
+    "wide-090",
+    "wide-091",
+    "wide-092",
+    "wide-093",
+    "wide-094",
+    "wide-095",
+    "wide-096",
+    "wide-097",
+    "wide-098",
+    "wide-099",
+    "wide-100",
+    "wide-101",
+    "wide-102",
+    "wide-103",
+    "wide-104",
+    "wide-105",
+    "wide-106",
+    "wide-107",
+    "wide-108",
+    "wide-109",
+    "wide-110",
+    "wide-111",
+    "wide-112",
+    "wide-113",
+    "wide-114",
+    "wide-115",
+    "wide-116",
+    "wide-117",
+    "wide-118",
+    "wide-119",
+    "wide-120",
+    "wide-121",
+    "wide-122",
+    "wide-123",
+    "wide-124",
+    "wide-125",
+    "wide-126",
+    "wide-127",
+    "wide-128",
+    "wide-129",
+    "wide-130",
+    "wide-131",
+    "wide-132",
+    "wide-133",
+    "wide-134",
+    "wide-135",
+    "wide-136",
+    "wide-137",
+    "wide-138",
+    "wide-139",
+    "wide-140",
+    "wide-141",
+    "wide-142",
+    "wide-143",
+    "wide-144",
+    "wide-145",
+    "wide-146",
+    "wide-147",
+    "wide-148",
+    "wide-149",
+    "wide-150",
+    "wide-151",
+    "wide-152",
+    "wide-153",
+    "wide-154",
+    "wide-155",
+    "wide-156",
+    "wide-157",
+    "wide-158",
+    "wide-159",
+    "wide-160",
+    "wide-161",
+    "wide-162",
+    "wide-163",
+    "wide-164",
+    "wide-165",
+    "wide-166",
+    "wide-167",
+    "wide-168",
+    "wide-169",
+    "wide-170",
+    "wide-171",
+    "wide-172",
+    "wide-173",
+    "wide-174",
+    "wide-175",
+    "wide-176",
+    "wide-177",
+    "wide-178",
+    "wide-179",
+    "wide-180",
+    "wide-181",
+    "wide-182",
+    "wide-183",
+    "wide-184",
+    "wide-185",
+    "wide-186",
+    "wide-187",
+    "wide-188",
+    "wide-189",
+    "wide-190",
+    "wide-191",
+    "wide-192",
+    "wide-193",
+    "wide-194",
+    "wide-195",
+    "wide-196",
+    "wide-197",
+    "wide-198",
+    "wide-199",
+    "wide-200",
+    "wide-201",
+    "wide-202",
+    "wide-203",
+    "wide-204",
+    "wide-205",
+    "wide-206",
+    "wide-207",
+    "wide-208",
+    "wide-209",
+    "wide-210",
+    "wide-211",
+    "wide-212",
+    "wide-213",
+    "wide-214",
+    "wide-215",
+    "wide-216",
+    "wide-217",
+    "wide-218",
+    "wide-219",
+    "wide-220",
+    "wide-221",
+    "wide-222",
+    "wide-223",
+    "wide-224",
+    "wide-225",
+    "wide-226",
+    "wide-227",
+    "wide-228",
+    "wide-229",
+    "wide-230",
+    "wide-231",
+    "wide-232",
+    "wide-233",
+    "wide-234",
+    "wide-235",
+    "wide-236",
+    "wide-237",
+    "wide-238",
+    "wide-239",
+    "wide-240",
+    "wide-241",
+    "wide-242",
+    "wide-243",
+    "wide-244",
+    "wide-245",
+    "wide-246",
+    "wide-247",
+    "wide-248",
+    "wide-249",
+    "wide-250",
+    "wide-251",
+    "wide-252",
+    "wide-253",
+    "wide-254",
+    "wide-255",
+    "wide-256",
+    "wide-257",
+    "wide-258",
+    "wide-259",
+    "wide-260",
+    "wide-261",
+    "wide-262",
+    "wide-263",
+    "wide-264",
+    "wide-265",
+    "wide-266",
+    "wide-267",
+    "wide-268",
+    "wide-269",
+    "wide-270",
+    "wide-271",
+    "wide-272",
+    "wide-273",
+    "wide-274",
+    "wide-275",
+    "wide-276",
+    "wide-277",
+    "wide-278",
+    "wide-279",
+    "wide-280",
+    "wide-281",
+    "wide-282",
+    "wide-283",
+    "wide-284",
+    "wide-285",
+    "wide-286",
+    "wide-287",
+    "wide-288",
+    "wide-289",
+    "wide-290",
+    "wide-291",
+    "wide-292",
+    "wide-293",
+    "wide-294",
+    "wide-295",
+    "wide-296",
+    "wide-297",
+    "wide-298",
+    "wide-299",
+    "wide-300",
+    "wide-301",
+    "wide-302",
+    "wide-303",
+    "wide-304",
+    "wide-305",
+    "wide-306",
+    "wide-307",
+    "wide-308",
+    "wide-309",
+    "wide-310",
+    "wide-311",
+    "wide-312",
+    "wide-313",
+    "wide-314",
+    "wide-315",
+    "wide-316",
+    "wide-317",
+    "wide-318",
+    "wide-319",
+    "wide-320"
+  ],
+  "selected_count": 320,
+  "selected_dev_overlap": 42,
+  "corrective_request_limit": 1,
+  "format_repair_limit": 1,
+  "repair_policy": "At most one total corrective request for format, citations or retrieval-context violations. Same question/context, no gold; retain all attempts.",
+  "dev_results_sha256": "4af9b52c250c837707c0690d892ec833347b3d3ee71d338ec631f6450eef8045",
+  "policy": "Evaluate every declared regression case, including DEV overlap; retain all failures and stop tuning after this freeze.",
+  "selected_categories": {
+    "shared_keyword_pair": 40,
+    "same_academic_year_pair": 40,
+    "same_start_year_pair": 32,
+    "person_title_year": 24,
+    "person_keyword_title": 32,
+    "document_author_category_multi_hop": 24,
+    "person_multiple_categories": 24,
+    "candidate_set_aggregation": 32,
+    "conjunctive_disambiguation": 40,
+    "negative_lookup": 32
+  },
+  "limitations": [
+    "All questions and previous model outputs are known; this is historical regression.",
+    "Harness supplies graph context; native OpenWebUI tool choice and browser UI are outside this experiment.",
+    "Checks structured factual claims, their source support/citations and required prose values; not a general semantic judge of arbitrary prose."
+  ]
+}

tools/graphrag/evaluation/results/answers_v7_test_2026-10-26/answers_test.csv

@@ -0,0 +1,81 @@
+id,category,strict,error,latency_seconds,attempts,first_pass_strict
+fresh7-001,shared_keyword_pair,True,,11.841625973989721,1,True
+fresh7-002,shared_keyword_pair,True,,8.292787773010787,1,True
+fresh7-003,shared_keyword_pair,True,,6.089656800992088,1,True
+fresh7-004,shared_keyword_pair,True,,7.950799065991305,1,True
+fresh7-005,shared_keyword_pair,True,,9.355126146008843,1,True
+fresh7-006,shared_keyword_pair,True,,8.119235402002232,1,True
+fresh7-007,shared_keyword_pair,True,,9.431924215008621,1,True
+fresh7-008,shared_keyword_pair,True,,8.421721135993721,1,True
+fresh7-009,shared_keyword_pair,True,,9.801845731999492,1,True
+fresh7-010,shared_keyword_pair,True,,8.407548419010709,1,True
+fresh7-011,same_academic_year_pair,True,,9.526154963998124,1,True
+fresh7-012,same_academic_year_pair,True,,7.739766710001277,1,True
+fresh7-013,same_academic_year_pair,True,,11.054396457999246,1,True
+fresh7-014,same_academic_year_pair,True,,5.981948126995121,1,True
+fresh7-015,same_academic_year_pair,True,,9.539100390000385,1,True
+fresh7-016,same_academic_year_pair,True,,6.883040939006605,1,True
+fresh7-017,same_academic_year_pair,True,,9.332280251008342,1,True
+fresh7-018,same_academic_year_pair,True,,6.666328727005748,1,True
+fresh7-019,same_academic_year_pair,True,,8.958197790008853,1,True
+fresh7-020,same_academic_year_pair,True,,7.030446254997514,1,True
+fresh7-021,same_start_year_pair,True,,9.740118704998167,1,True
+fresh7-022,same_start_year_pair,True,,7.7155549000017345,1,True
+fresh7-023,same_start_year_pair,True,,9.746938041003887,1,True
+fresh7-024,same_start_year_pair,True,,7.16414850000001,1,True
+fresh7-025,same_start_year_pair,True,,9.141140612002346,1,True
+fresh7-026,same_start_year_pair,True,,6.95191725899349,1,True
+fresh7-027,same_start_year_pair,True,,9.19188570100232,1,True
+fresh7-028,same_start_year_pair,True,,7.348483943002066,1,True
+fresh7-029,same_start_year_pair,True,,8.995747175998986,1,True
+fresh7-030,same_start_year_pair,True,,6.975118151007337,1,True
+fresh7-031,person_title_year,True,,7.311062619992299,1,True
+fresh7-032,person_title_year,True,,5.878452272008872,1,True
+fresh7-033,person_title_year,True,,5.801723632990615,1,True
+fresh7-034,person_title_year,True,,5.779961151012685,1,True
+fresh7-035,person_title_year,True,,5.503505647997372,1,True
+fresh7-036,person_title_year,True,,5.739178684991202,1,True
+fresh7-037,person_title_year,True,,5.415736306007602,1,True
+fresh7-038,person_title_year,True,,7.437335399008589,1,True
+fresh7-039,person_title_year,True,,4.408995279998635,1,True
+fresh7-040,person_title_year,True,,6.882196734994068,1,True
+fresh7-041,person_keyword_title,True,,5.83978758000012,1,True
+fresh7-042,person_keyword_title,True,,8.581415363005362,1,True
+fresh7-043,person_keyword_title,True,,7.124095003004186,1,True
+fresh7-044,person_keyword_title,True,,9.642784591007512,1,True
+fresh7-045,person_keyword_title,True,,7.515784107992658,1,True
+fresh7-046,person_keyword_title,True,,10.188013713996042,1,True
+fresh7-047,person_keyword_title,True,,8.431851733999792,1,True
+fresh7-048,person_keyword_title,True,,9.907063878999907,1,True
+fresh7-049,person_keyword_title,True,,8.099193093003123,1,True
+fresh7-050,person_keyword_title,True,,8.852308599001844,1,True
+fresh7-051,document_author_category_multi_hop,True,,7.668113538005855,1,True
+fresh7-052,document_author_category_multi_hop,True,,8.831390542007284,1,True
+fresh7-053,person_multiple_categories,True,,4.3355173820018535,1,True
+fresh7-054,person_multiple_categories,True,,14.959355758997845,2,False
+fresh7-055,person_multiple_categories,True,,4.788777031993959,1,True
+fresh7-056,person_multiple_categories,True,,6.780442283998127,1,True
+fresh7-057,person_multiple_categories,True,,5.894032968004467,1,True
+fresh7-058,person_multiple_categories,True,,16.16342897800496,2,False
+fresh7-059,person_multiple_categories,True,,8.475595613010228,2,False
+fresh7-060,candidate_set_aggregation,True,,9.091813248000108,1,True
+fresh7-061,candidate_set_aggregation,True,,8.758890958008124,1,True
+fresh7-062,candidate_set_aggregation,True,,8.99064658000134,1,True
+fresh7-063,candidate_set_aggregation,True,,9.600775625003735,1,True
+fresh7-064,candidate_set_aggregation,True,,8.849686492001638,1,True
+fresh7-065,candidate_set_aggregation,True,,9.296292756989715,1,True
+fresh7-066,candidate_set_aggregation,True,,8.94247531799192,1,True
+fresh7-067,conjunctive_disambiguation,True,,8.469535019001341,1,True
+fresh7-068,conjunctive_disambiguation,True,,8.207272173007368,1,True
+fresh7-069,conjunctive_disambiguation,True,,7.925748694004142,1,True
+fresh7-070,conjunctive_disambiguation,True,,8.149811054012389,1,True
+fresh7-071,conjunctive_disambiguation,True,,8.003814060997684,1,True
+fresh7-072,conjunctive_disambiguation,True,,8.130394503008574,1,True
+fresh7-073,negative_lookup,True,,4.625077007003711,1,True
+fresh7-074,negative_lookup,True,,4.179091968006105,1,True
+fresh7-075,negative_lookup,True,,3.3433215739933075,1,True
+fresh7-076,negative_lookup,True,,3.1084746859996812,1,True
+fresh7-077,negative_lookup,True,,2.6855998700048076,1,True
+fresh7-078,negative_lookup,True,,2.4824905980058247,1,True
+fresh7-079,negative_lookup,True,,2.805210263992194,1,True
+fresh7-080,negative_lookup,True,,1.868269345999579,1,True

tools/graphrag/evaluation/results/answers_v7_test_2026-10-26/frozen_answers.json

@@ -0,0 +1,259 @@
+{
+  "method": "Fresh query-group holdout over known sources and templates; zero DEV query-group overlap.",
+  "graph_freeze_sha256": "9e97794200520d0f2288e452c9e186c0291a54c63d2c1382f693a9753e4081cf",
+  "audit_sha256": "7825d4f3e4e535eb5f49e975dfef363b9b813f1de5c97e2387e5c2f3e578e6a5",
+  "generation_files": {
+    "tools/answer_evaluation/evaluate_graph_answers.py": "d81504e8fe51598c6bd1b0d482f61b9b6f51f87ba3f38b03ae1bc584675a3b76",
+    "tools/answer_evaluation/graph_answer_prompt.txt": "fc9f2f8d52e70a8d67b9f4814b3f15e6be64bf88afe53ce4db4f037bde00e064",
+    "tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
+    "tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
+    "tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
+    "tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002",
+    "tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1"
+  },
+  "model": "model120-fast",
+  "temperature": 0,
+  "workers": 3,
+  "dev_ids": [
+    "v7-dev-year-1899",
+    "v7-dev-year-1999",
+    "v7-dev-year-2100",
+    "v7-dev-year-9999",
+    "wide-001",
+    "wide-002",
+    "wide-041",
+    "wide-042",
+    "wide-080",
+    "wide-081",
+    "wide-082",
+    "wide-104",
+    "wide-113",
+    "wide-114",
+    "wide-116",
+    "wide-119",
+    "wide-129",
+    "wide-137",
+    "wide-138",
+    "wide-139",
+    "wide-141",
+    "wide-142",
+    "wide-143",
+    "wide-145",
+    "wide-155",
+    "wide-159",
+    "wide-161",
+    "wide-163",
+    "wide-165",
+    "wide-169",
+    "wide-170",
+    "wide-178",
+    "wide-181",
+    "wide-193",
+    "wide-194",
+    "wide-209",
+    "wide-217",
+    "wide-218",
+    "wide-232",
+    "wide-236",
+    "wide-239",
+    "wide-249",
+    "wide-250",
+    "wide-283",
+    "wide-289",
+    "wide-290"
+  ],
+  "test_ids": [
+    "fresh7-001",
+    "fresh7-002",
+    "fresh7-003",
+    "fresh7-004",
+    "fresh7-005",
+    "fresh7-006",
+    "fresh7-007",
+    "fresh7-008",
+    "fresh7-009",
+    "fresh7-010",
+    "fresh7-011",
+    "fresh7-012",
+    "fresh7-013",
+    "fresh7-014",
+    "fresh7-015",
+    "fresh7-016",
+    "fresh7-017",
+    "fresh7-018",
+    "fresh7-019",
+    "fresh7-020",
+    "fresh7-021",
+    "fresh7-022",
+    "fresh7-023",
+    "fresh7-024",
+    "fresh7-025",
+    "fresh7-026",
+    "fresh7-027",
+    "fresh7-028",
+    "fresh7-029",
+    "fresh7-030",
+    "fresh7-031",
+    "fresh7-032",
+    "fresh7-033",
+    "fresh7-034",
+    "fresh7-035",
+    "fresh7-036",
+    "fresh7-037",
+    "fresh7-038",
+    "fresh7-039",
+    "fresh7-040",
+    "fresh7-041",
+    "fresh7-042",
+    "fresh7-043",
+    "fresh7-044",
+    "fresh7-045",
+    "fresh7-046",
+    "fresh7-047",
+    "fresh7-048",
+    "fresh7-049",
+    "fresh7-050",
+    "fresh7-051",
+    "fresh7-052",
+    "fresh7-053",
+    "fresh7-054",
+    "fresh7-055",
+    "fresh7-056",
+    "fresh7-057",
+    "fresh7-058",
+    "fresh7-059",
+    "fresh7-060",
+    "fresh7-061",
+    "fresh7-062",
+    "fresh7-063",
+    "fresh7-064",
+    "fresh7-065",
+    "fresh7-066",
+    "fresh7-067",
+    "fresh7-068",
+    "fresh7-069",
+    "fresh7-070",
+    "fresh7-071",
+    "fresh7-072",
+    "fresh7-073",
+    "fresh7-074",
+    "fresh7-075",
+    "fresh7-076",
+    "fresh7-077",
+    "fresh7-078",
+    "fresh7-079",
+    "fresh7-080"
+  ],
+  "dev_count": 46,
+  "test_count": 80,
+  "target_accuracy": 0.92,
+  "evaluation_kind": "independent_query_group_holdout",
+  "rendering": "Verified structured claims rendered deterministically; raw model outputs preserved.",
+  "selected_split": "test",
+  "selected_ids": [
+    "fresh7-001",
+    "fresh7-002",
+    "fresh7-003",
+    "fresh7-004",
+    "fresh7-005",
+    "fresh7-006",
+    "fresh7-007",
+    "fresh7-008",
+    "fresh7-009",
+    "fresh7-010",
+    "fresh7-011",
+    "fresh7-012",
+    "fresh7-013",
+    "fresh7-014",
+    "fresh7-015",
+    "fresh7-016",
+    "fresh7-017",
+    "fresh7-018",
+    "fresh7-019",
+    "fresh7-020",
+    "fresh7-021",
+    "fresh7-022",
+    "fresh7-023",
+    "fresh7-024",
+    "fresh7-025",
+    "fresh7-026",
+    "fresh7-027",
+    "fresh7-028",
+    "fresh7-029",
+    "fresh7-030",
+    "fresh7-031",
+    "fresh7-032",
+    "fresh7-033",
+    "fresh7-034",
+    "fresh7-035",
+    "fresh7-036",
+    "fresh7-037",
+    "fresh7-038",
+    "fresh7-039",
+    "fresh7-040",
+    "fresh7-041",
+    "fresh7-042",
+    "fresh7-043",
+    "fresh7-044",
+    "fresh7-045",
+    "fresh7-046",
+    "fresh7-047",
+    "fresh7-048",
+    "fresh7-049",
+    "fresh7-050",
+    "fresh7-051",
+    "fresh7-052",
+    "fresh7-053",
+    "fresh7-054",
+    "fresh7-055",
+    "fresh7-056",
+    "fresh7-057",
+    "fresh7-058",
+    "fresh7-059",
+    "fresh7-060",
+    "fresh7-061",
+    "fresh7-062",
+    "fresh7-063",
+    "fresh7-064",
+    "fresh7-065",
+    "fresh7-066",
+    "fresh7-067",
+    "fresh7-068",
+    "fresh7-069",
+    "fresh7-070",
+    "fresh7-071",
+    "fresh7-072",
+    "fresh7-073",
+    "fresh7-074",
+    "fresh7-075",
+    "fresh7-076",
+    "fresh7-077",
+    "fresh7-078",
+    "fresh7-079",
+    "fresh7-080"
+  ],
+  "selected_count": 80,
+  "selected_dev_overlap": 0,
+  "corrective_request_limit": 1,
+  "format_repair_limit": 1,
+  "repair_policy": "At most one total corrective request for format, citations or retrieval-context violations. Same question/context, no gold; retain all attempts.",
+  "dev_results_sha256": "4af9b52c250c837707c0690d892ec833347b3d3ee71d338ec631f6450eef8045",
+  "policy": "Evaluate every declared fresh TEST case once; retain every failure; do not tune on this TEST.",
+  "selected_categories": {
+    "shared_keyword_pair": 10,
+    "same_academic_year_pair": 10,
+    "same_start_year_pair": 10,
+    "person_title_year": 10,
+    "person_keyword_title": 10,
+    "document_author_category_multi_hop": 2,
+    "person_multiple_categories": 7,
+    "candidate_set_aggregation": 7,
+    "conjunctive_disambiguation": 6,
+    "negative_lookup": 8
+  },
+  "limitations": [
+    "Known corpus/profiles and templates; these query groups were not previously evaluated.",
+    "Harness supplies graph context; native OpenWebUI tool choice and browser UI are outside this experiment.",
+    "Checks structured factual claims, their source support/citations and required prose values; not a general semantic judge of arbitrary prose."
+  ]
+}

tools/graphrag/evaluation/results/final_v7_2026-10-26/REPORT.md

@@ -0,0 +1,147 @@
+# GraphRAG v7 — uzavretie bodu 3, 26. 10. 2026
+
+Nový query-group TEST dosiahol **80/80** pri grafovom vyhľadávaní aj pri
+finálnych odpovediach. Na prvý pokus modelu prešlo **77/80 — 96,25 %**;
+tri otázky prešli po jednom povolenom opravnom požiadavku. Cieľ ≥92 % je splnený
+už pri prvom pokuse aj pri finálnej pipeline. Celá sada má **591 úspešných
+testov, bez preskočenia**.
+
+**Bod 3 je technicky uzavretý pre podporované typy dotazov a tento korpus.**
+Hodnotí sa modelový výber štruktúrovaných tvrdení, kontrola zdrojov a
+renderer; nie ľubovoľná voľná próza LLM ani natívne OpenWebUI smerovanie.
+100 % na 80 otázkach nie je zárukou všeobecnej pravdivosti.
+
+## Výsledky
+
+| Meranie | Správne | Úspešnosť |
+|---|---:|---:|
+| Graf DEV vrátane známych regresií | 1 046/1 046 | 100,00 % |
+| Graf nový query-group TEST | 80/80 | 100,00 % |
+| Graf historická podmnožina v DEV | 320/320 | 100,00 % |
+| Model + renderer DEV, finálne | 46/46 | 100,00 % |
+| Nový TEST, prvý pokus modelu + renderer | 77/80 | 96,25 % |
+| Nový TEST, po jednej povolenej oprave | 80/80 | 100,00 % |
+| Historická regresia, prvý pokus | 310/320 | 96.88 % |
+| Historická regresia, finálne | 319/320 | 99.69 % |
+| Živé syntetické funkčné scenáre mimo benchmarku | 9/9 | 100,00 % |
+
+Modelový DEV mal prvý pokus 45/46 a jednu úspešnú opravu formátu.
+V novom TESTe potreboval `fresh7-054` opravu schémy a `fresh7-058`,
+`fresh7-059` odstránenie nepodloženého počtu pri neagregačnej otázke.
+Opravu vyvolal iba formát alebo skutočný kontext; gold ani následné
+skóre sa pri rozhodovaní o opakovaní nepoužívajú.
+Finálna historická regresia má **0 chýb spracovania**,
+**10 opravných požiadavkov**
+(z toho formát 3, kontext 7).
+Všetky odpovede, pôvodné pokusy a prípadné zlyhania ostávajú v menovateli.
+
+Zostávajúci prípad `wide-285`: po chybe schémy model vložil názov dcérskeho
+dokumentu PYTORCH - LSTM TUTORIÁL ako akademický názov k roku 2019.
+Taká väzba nie je v akademických faktoch doložená; guard ju odmietol
+(`unsupported_academic_binding`, `accepted_for_delivery=false`).
+Odpoveď sa nedoručila. Povolený opravný pokus už bol vyčerpaný.
+[Úplný záznam](../answers_v7_regression_2026-10-26/failures.json).
+
+| Kategória nového TESTu | Finálne správne | Úspešnosť |
+|---|---:|---:|
+| `candidate_set_aggregation` | 7/7 | 100.00 % |
+| `conjunctive_disambiguation` | 6/6 | 100.00 % |
+| `document_author_category_multi_hop` | 2/2 | 100.00 % |
+| `negative_lookup` | 8/8 | 100.00 % |
+| `person_keyword_title` | 10/10 | 100.00 % |
+| `person_multiple_categories` | 7/7 | 100.00 % |
+| `person_title_year` | 10/10 | 100.00 % |
+| `same_academic_year_pair` | 10/10 | 100.00 % |
+| `same_start_year_pair` | 10/10 | 100.00 % |
+| `shared_keyword_pair` | 10/10 | 100.00 % |
+
+## Uzavreté podbody
+
+| Podbod | Konkrétny výsledok |
+|---|---|
+| 3.1 Benchmark a zdroje | 1 126 zdrojovo skontrolovaných otázok; 10 opravených očakávaní po odstránení odhadu štúdia z adresára. Nová vzorka 20 otázok priamo skontrolovaná z Markdownov. |
+| 3.2 Evaluator a doručenie | Rozpory v úplnosti, nepodložené počty/roky, chýbajúce claims/citácie a HTML/referenčné odkazy majú regresné testy. Finálny text vzniká iba z overených tvrdení. |
+| 3.3 DEV | Kompletný grafový DEV 1 046/1 046, modelový DEV 46/46; staré prípady `wide-104`, `wide-119`, `wide-143`, `wide-176` prešli na prvý pokus aj v novej regresii. |
+| 3.4 Zmrazenie | `frozen_v7.json`, deklarácie generovania a `frozen_release_v7.json`; explicitné CLI vstupy, odtlačky rendereru aj spoločných modulov. |
+| 3.5 Finálne overenie | Nový TEST 80 skupín bez DEV prekryvu, jednorazovo po zmrazení; samostatná historická regresia 320 otázok a kontrola Dockeru. |
+| 3.6 Metriky a správa | Oddelené prvotné a finálne odpovede, vyhľadávanie a modelová pipeline, nový TEST a regresia. V1–v6 ostávajú zachované. |
+
+## Opravené chyby
+
+- Rok 1999 sa už nevynechá: dotaz na Jána Holpa nevracia práce z 2019/2021.
+  Aj ostatné štvormiestne roky sa aplikujú ako obmedzenie. Nepodporovaný
+  číselný zápis po „roku“ sa explicitne odmieta.
+- Veta „Výpis je neúplný“ neospravedlňuje iné kladné tvrdenie o úplnom zozname.
+  Pravdivé popretie úplnosti je povolené.
+- Prirodzený text sa skladá deterministicky z validovaných claims/count a URL.
+  Pri regrade neočakávanej voľnej prózy dopísaný rok 2099 neprejde.
+  Zostávajúci číselný filter nie je univerzálny sémantický analyzátor;
+  používateľovi sa preto nedoručuje nekontrolovaný voľný text modelu.
+- Počet bez úplnej agregácie sa odmietne. Pozitívne faktové odpovede vyžadujú
+  tvrdenia a citácie, vrátane zdrojov jednotlivých atribútov. Count=0 v úplnej
+  prázdnej agregácii je povolený.
+- Kontrolujú sa aj referenčné Markdown a HTML odkazy, relatívne URL, entity
+  a korektné URL so zátvorkami. Nepovolené odkazy sa neobchádzajú upozornením.
+- Začiatok štúdia je iba explicitne uvedený údaj. Graf eviduje pôvod a istotu,
+  chýbajúca hodnota ostáva neznáma; `wide-112` už nevyžaduje odhadovaný rok
+  Tomáša Kuchárika. Cohort nie je administratívny fakt o nástupe na štúdium.
+- Skloňované meno sa v texte rozpoznáva obmedzenými príponami pri zachovaní
+  presnej štruktúrovanej identity. Canonical renderer používa pôvodné meno.
+- Dataset, audit, manifest a DEV ID sa pri evaluácii vyberajú explicitne;
+  spúšťanie už potichu nevyberá v4.
+
+## Nezávislosť a pôvod očakávaní
+
+Všetkých 1 042 v6 otázok je odhalených a v7 ich označuje ako DEV. Štyri
+nové auditové rokové regresie dopĺňajú počet na 1 046. Historických 320
+sa vyhodnocuje z tejto množiny samostatne. Je to **regresia**, nie nový TEST.
+
+Nový TEST má 80 kombinácií, 72 kladných a 8 záporných. Seed 20261026,
+kvóty a ID vznikli z raw Markdownov pred čítaním predikcií. Normalizované
+query skupiny ani formulácie sa s DEV neprekrývajú. Po zmrazení sa podľa
+TESTu nemenil kód, prompt, gold ani výber otázok. Korpus, osoby a rodiny
+šablón sú známe; nejde o unseen-profile experiment.
+
+Priama AI kontrola 20 otázok pokrýva všetkých desať kategórií.
+[Zdrojový hárok](../v7_preparation/source_review.csv) a
+[výňatky](../v7_preparation/source_review_packet.json) sú uložené.
+**Ľudská anotácia neprebehla.** Neistá adresárová konvencia bola odstránená,
+nie potvrdená bez dôkazu. Autor–kategória má iba dve nové skupiny a výpis
+kategórií sedem; nemožno tvrdiť rovnakú štatistickú silu každej kategórie.
+
+## Regresia a zachované referencie
+
+V6 mala strict 316/320. V7 mení gold pri 10 známych otázkach (sedem z
+historických 320) a používa štruktúrovaný modelový výstup s rendererom.
+Rozdiel percent nie je izolovaný odhad zlepšenia samotného LLM.
+Všetky pôvodné baseline, otázky, finálne JSON/CSV a v1–v6 výsledky sa zachovávajú.
+
+Prepočet citácií 56 archivovaných súborov má 6 044 odpovedí a päť výskytov
+URL mimo podkladov, rovnako ako v6. Ide o prekrývajúce sa experimenty.
+Klasický RAG sa novým modelovým TESTom nemeral. Prehratie piatich uložených
+RAG odpovedí cez kontrolu upozornenia opäť prešlo 5/5; to nie je nová
+presnosť modelu. Je to regresia prezentačnej vrstvy.
+
+## Nasadenie a reprodukcia
+
+Súbory, SQLite a Neo4j sú po koordinovanom `/sync` na spoločnej generácii
+`33ea0794472c416e97c84371a2a6cdf8`; wiki Git je
+`3e892a1d49e9bce118cd03478ce15e01c2bebb4e`. Index má 114 dokumentov a 887
+chunkov, graf 446 uzlov a 877 vzťahov. Explicitné published=false sa skryje,
+16 chýbajúcich hodnôt sa zahrnie; viditeľných je 113 dokumentov.
+
+[Hostiteľské overenie](deployment_check_host.json) a
+[overenie Dockeru](deployment_check_docker.json) potvrdzujú dostupnosť,
+spoločnú generáciu, normalizované publikovanie a runtime odtlačky.
+Modelové výstupy obsahujú celý skutočný grafový kontext a verziu dát.
+
+[Príkazy a zmeny](../../../../../docs/graphrag_v7_2026-10-26_worklog.md).
+Reprodukcia používa nový výstupný adresár a príslušný release checkout.
+Git HEAD v metadátach merania je východiskový commit pred finálnym commitom;
+konkrétny meraný kód určujú obsahové SHA. Model `model120-fast` je vzdialená
+služba; parametre a odpovede sú uložené, váhy prevádzkuje poskytovateľ.
+
+Ďalším blokom sú natívne nástroje a hlavný agent v OpenWebUI. Musia rešpektovať
+odmietnutie doručenia a odovzdať overený text; voľné prepísanie rendererového
+textu bez kontrol by uvedené vlastnosti obišlo. Toto smerovanie sa v bode 3
+nemeralo a je súčasťou bodu 4.

tools/graphrag/evaluation/results/final_v7_2026-10-26/deployment_check_docker.json

@@ -0,0 +1,130 @@
+{
+  "checked_at_utc": "2026-10-26T00:38:33.284887+00:00",
+  "health": {
+    "status": "ok",
+    "ready": true,
+    "database_exists": true,
+    "search_engine": "hybrid_fts5_embeddings",
+    "rag_enabled": true,
+    "zpwiki_exists": true,
+    "security_configured": true,
+    "index_version": {
+      "version": "33ea0794472c416e97c84371a2a6cdf8",
+      "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+      "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+      "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+      "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+      "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+      "embedding_model": "intfloat/multilingual-e5-small",
+      "publication_policy": "explicit-false-hidden-v1"
+    },
+    "tools": {
+      "rag": {
+        "ready": true,
+        "status": "ready",
+        "reason": null,
+        "embedding_model": "intfloat/multilingual-e5-small",
+        "dimensions": 384,
+        "index_version": {
+          "version": "33ea0794472c416e97c84371a2a6cdf8",
+          "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+          "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+          "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+          "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+          "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+          "embedding_model": "intfloat/multilingual-e5-small",
+          "publication_policy": "explicit-false-hidden-v1"
+        }
+      },
+      "neo4j": {
+        "ready": true,
+        "status": "ready",
+        "reason": null
+      },
+      "graphrag": {
+        "ready": true,
+        "status": "ready",
+        "reason": null,
+        "index_version": {
+          "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+          "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+          "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+          "embedding_model": "intfloat/multilingual-e5-small",
+          "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+          "publication_policy": "explicit-false-hidden-v1",
+          "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+          "version": "33ea0794472c416e97c84371a2a6cdf8"
+        },
+        "nodes": 446
+      }
+    }
+  },
+  "index_version": {
+    "version": "33ea0794472c416e97c84371a2a6cdf8",
+    "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+    "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+    "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+    "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+    "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+    "embedding_model": "intfloat/multilingual-e5-small",
+    "publication_policy": "explicit-false-hidden-v1"
+  },
+  "frozen_implementation_inputs_and_graph_match": true,
+  "graph": {
+    "sha256": "566817ce7ddebb2638585192c148c41421071bdf719674fc6ddddb1c99a0e36f",
+    "nodes": 446,
+    "relationships": 877
+  },
+  "graph_cases": {
+    "known_title": {
+      "status": "ok",
+      "reason": null,
+      "complete": true
+    },
+    "generic_label_not_title": {
+      "status": "no_answer",
+      "reason": "no_matching_facts",
+      "complete": true
+    },
+    "unknown_person": {
+      "status": "no_answer",
+      "reason": "unknown_person",
+      "complete": true
+    },
+    "outside_century": {
+      "status": "no_answer",
+      "reason": "no_matching_facts",
+      "complete": true
+    },
+    "unsupported_year": {
+      "status": "no_answer",
+      "reason": "unsupported_year_constraint",
+      "complete": true
+    },
+    "unknown_start": {
+      "status": "no_answer",
+      "reason": "missing_documented_start_year",
+      "complete": true
+    }
+  },
+  "rag_live_query": "ok",
+  "release_code_match": true,
+  "rag_context_metadata_matches_contract": true,
+  "publication": {
+    "documents": 114,
+    "explicit_false": 1,
+    "missing_included": 16,
+    "visible_by_policy": 113,
+    "hidden_document_absent_from_rag": true
+  },
+  "runtime": {
+    "python": "3.11.17",
+    "fastapi": "0.136.3",
+    "pydantic": "2.13.4",
+    "neo4j": "6.3.1",
+    "sentence-transformers": "5.7.0",
+    "torch": "2.11.0+cpu"
+  },
+  "scope": "RAG HTTP API plus GraphRAG library/CLI; native OpenWebUI graph-tool routing is point 4.",
+  "docker_image_id": "sha256:10fd5e24b91cf2907a2b712ac8c2af796a26f5023daaf3834f3ca505c62f36f6"
+}

tools/graphrag/evaluation/results/final_v7_2026-10-26/deployment_check_host.json

@@ -0,0 +1,129 @@
+{
+  "checked_at_utc": "2026-10-26T00:38:25.612948+00:00",
+  "health": {
+    "status": "ok",
+    "ready": true,
+    "database_exists": true,
+    "search_engine": "hybrid_fts5_embeddings",
+    "rag_enabled": true,
+    "zpwiki_exists": true,
+    "security_configured": true,
+    "index_version": {
+      "version": "33ea0794472c416e97c84371a2a6cdf8",
+      "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+      "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+      "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+      "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+      "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+      "embedding_model": "intfloat/multilingual-e5-small",
+      "publication_policy": "explicit-false-hidden-v1"
+    },
+    "tools": {
+      "rag": {
+        "ready": true,
+        "status": "ready",
+        "reason": null,
+        "embedding_model": "intfloat/multilingual-e5-small",
+        "dimensions": 384,
+        "index_version": {
+          "version": "33ea0794472c416e97c84371a2a6cdf8",
+          "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+          "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+          "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+          "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+          "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+          "embedding_model": "intfloat/multilingual-e5-small",
+          "publication_policy": "explicit-false-hidden-v1"
+        }
+      },
+      "neo4j": {
+        "ready": true,
+        "status": "ready",
+        "reason": null
+      },
+      "graphrag": {
+        "ready": true,
+        "status": "ready",
+        "reason": null,
+        "index_version": {
+          "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+          "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+          "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+          "embedding_model": "intfloat/multilingual-e5-small",
+          "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+          "publication_policy": "explicit-false-hidden-v1",
+          "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+          "version": "33ea0794472c416e97c84371a2a6cdf8"
+        },
+        "nodes": 446
+      }
+    }
+  },
+  "index_version": {
+    "version": "33ea0794472c416e97c84371a2a6cdf8",
+    "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+    "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+    "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+    "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+    "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+    "embedding_model": "intfloat/multilingual-e5-small",
+    "publication_policy": "explicit-false-hidden-v1"
+  },
+  "frozen_implementation_inputs_and_graph_match": true,
+  "graph": {
+    "sha256": "566817ce7ddebb2638585192c148c41421071bdf719674fc6ddddb1c99a0e36f",
+    "nodes": 446,
+    "relationships": 877
+  },
+  "graph_cases": {
+    "known_title": {
+      "status": "ok",
+      "reason": null,
+      "complete": true
+    },
+    "generic_label_not_title": {
+      "status": "no_answer",
+      "reason": "no_matching_facts",
+      "complete": true
+    },
+    "unknown_person": {
+      "status": "no_answer",
+      "reason": "unknown_person",
+      "complete": true
+    },
+    "outside_century": {
+      "status": "no_answer",
+      "reason": "no_matching_facts",
+      "complete": true
+    },
+    "unsupported_year": {
+      "status": "no_answer",
+      "reason": "unsupported_year_constraint",
+      "complete": true
+    },
+    "unknown_start": {
+      "status": "no_answer",
+      "reason": "missing_documented_start_year",
+      "complete": true
+    }
+  },
+  "rag_live_query": "ok",
+  "release_code_match": true,
+  "rag_context_metadata_matches_contract": true,
+  "publication": {
+    "documents": 114,
+    "explicit_false": 1,
+    "missing_included": 16,
+    "visible_by_policy": 113,
+    "hidden_document_absent_from_rag": true
+  },
+  "runtime": {
+    "python": "3.10.12",
+    "fastapi": "0.136.3",
+    "pydantic": "2.13.4",
+    "neo4j": "6.3.1",
+    "sentence-transformers": "5.7.0",
+    "torch": "2.11.0+cpu"
+  },
+  "scope": "RAG HTTP API plus GraphRAG library/CLI; native OpenWebUI graph-tool routing is point 4."
+}

tools/graphrag/evaluation/results/final_v7_2026-10-26/historical_graph_summary.json

@@ -0,0 +1,348 @@
+{
+  "method": "Historical 320-case subset of the complete 1046-case DEV; no duplicated retrieval run.",
+  "source": "tools/graphrag/evaluation/results/final_v7_2026-10-26/graphrag_dev.json",
+  "source_sha256": "d1a5d7b6d5aff3c0002f36a60cebc07627a79ec6d57de119dd587e8ab70143b0",
+  "ids": [
+    "wide-001",
+    "wide-002",
+    "wide-003",
+    "wide-004",
+    "wide-005",
+    "wide-006",
+    "wide-007",
+    "wide-008",
+    "wide-009",
+    "wide-010",
+    "wide-011",
+    "wide-012",
+    "wide-013",
+    "wide-014",
+    "wide-015",
+    "wide-016",
+    "wide-017",
+    "wide-018",
+    "wide-019",
+    "wide-020",
+    "wide-021",
+    "wide-022",
+    "wide-023",
+    "wide-024",
+    "wide-025",
+    "wide-026",
+    "wide-027",
+    "wide-028",
+    "wide-029",
+    "wide-030",
+    "wide-031",
+    "wide-032",
+    "wide-033",
+    "wide-034",
+    "wide-035",
+    "wide-036",
+    "wide-037",
+    "wide-038",
+    "wide-039",
+    "wide-040",
+    "wide-041",
+    "wide-042",
+    "wide-043",
+    "wide-044",
+    "wide-045",
+    "wide-046",
+    "wide-047",
+    "wide-048",
+    "wide-049",
+    "wide-050",
+    "wide-051",
+    "wide-052",
+    "wide-053",
+    "wide-054",
+    "wide-055",
+    "wide-056",
+    "wide-057",
+    "wide-058",
+    "wide-059",
+    "wide-060",
+    "wide-061",
+    "wide-062",
+    "wide-063",
+    "wide-064",
+    "wide-065",
+    "wide-066",
+    "wide-067",
+    "wide-068",
+    "wide-069",
+    "wide-070",
+    "wide-071",
+    "wide-072",
+    "wide-073",
+    "wide-074",
+    "wide-075",
+    "wide-076",
+    "wide-077",
+    "wide-078",
+    "wide-079",
+    "wide-080",
+    "wide-081",
+    "wide-082",
+    "wide-083",
+    "wide-084",
+    "wide-085",
+    "wide-086",
+    "wide-087",
+    "wide-088",
+    "wide-089",
+    "wide-090",
+    "wide-091",
+    "wide-092",
+    "wide-093",
+    "wide-094",
+    "wide-095",
+    "wide-096",
+    "wide-097",
+    "wide-098",
+    "wide-099",
+    "wide-100",
+    "wide-101",
+    "wide-102",
+    "wide-103",
+    "wide-104",
+    "wide-105",
+    "wide-106",
+    "wide-107",
+    "wide-108",
+    "wide-109",
+    "wide-110",
+    "wide-111",
+    "wide-112",
+    "wide-113",
+    "wide-114",
+    "wide-115",
+    "wide-116",
+    "wide-117",
+    "wide-118",
+    "wide-119",
+    "wide-120",
+    "wide-121",
+    "wide-122",
+    "wide-123",
+    "wide-124",
+    "wide-125",
+    "wide-126",
+    "wide-127",
+    "wide-128",
+    "wide-129",
+    "wide-130",
+    "wide-131",
+    "wide-132",
+    "wide-133",
+    "wide-134",
+    "wide-135",
+    "wide-136",
+    "wide-137",
+    "wide-138",
+    "wide-139",
+    "wide-140",
+    "wide-141",
+    "wide-142",
+    "wide-143",
+    "wide-144",
+    "wide-145",
+    "wide-146",
+    "wide-147",
+    "wide-148",
+    "wide-149",
+    "wide-150",
+    "wide-151",
+    "wide-152",
+    "wide-153",
+    "wide-154",
+    "wide-155",
+    "wide-156",
+    "wide-157",
+    "wide-158",
+    "wide-159",
+    "wide-160",
+    "wide-161",
+    "wide-162",
+    "wide-163",
+    "wide-164",
+    "wide-165",
+    "wide-166",
+    "wide-167",
+    "wide-168",
+    "wide-169",
+    "wide-170",
+    "wide-171",
+    "wide-172",
+    "wide-173",
+    "wide-174",
+    "wide-175",
+    "wide-176",
+    "wide-177",
+    "wide-178",
+    "wide-179",
+    "wide-180",
+    "wide-181",
+    "wide-182",
+    "wide-183",
+    "wide-184",
+    "wide-185",
+    "wide-186",
+    "wide-187",
+    "wide-188",
+    "wide-189",
+    "wide-190",
+    "wide-191",
+    "wide-192",
+    "wide-193",
+    "wide-194",
+    "wide-195",
+    "wide-196",
+    "wide-197",
+    "wide-198",
+    "wide-199",
+    "wide-200",
+    "wide-201",
+    "wide-202",
+    "wide-203",
+    "wide-204",
+    "wide-205",
+    "wide-206",
+    "wide-207",
+    "wide-208",
+    "wide-209",
+    "wide-210",
+    "wide-211",
+    "wide-212",
+    "wide-213",
+    "wide-214",
+    "wide-215",
+    "wide-216",
+    "wide-217",
+    "wide-218",
+    "wide-219",
+    "wide-220",
+    "wide-221",
+    "wide-222",
+    "wide-223",
+    "wide-224",
+    "wide-225",
+    "wide-226",
+    "wide-227",
+    "wide-228",
+    "wide-229",
+    "wide-230",
+    "wide-231",
+    "wide-232",
+    "wide-233",
+    "wide-234",
+    "wide-235",
+    "wide-236",
+    "wide-237",
+    "wide-238",
+    "wide-239",
+    "wide-240",
+    "wide-241",
+    "wide-242",
+    "wide-243",
+    "wide-244",
+    "wide-245",
+    "wide-246",
+    "wide-247",
+    "wide-248",
+    "wide-249",
+    "wide-250",
+    "wide-251",
+    "wide-252",
+    "wide-253",
+    "wide-254",
+    "wide-255",
+    "wide-256",
+    "wide-257",
+    "wide-258",
+    "wide-259",
+    "wide-260",
+    "wide-261",
+    "wide-262",
+    "wide-263",
+    "wide-264",
+    "wide-265",
+    "wide-266",
+    "wide-267",
+    "wide-268",
+    "wide-269",
+    "wide-270",
+    "wide-271",
+    "wide-272",
+    "wide-273",
+    "wide-274",
+    "wide-275",
+    "wide-276",
+    "wide-277",
+    "wide-278",
+    "wide-279",
+    "wide-280",
+    "wide-281",
+    "wide-282",
+    "wide-283",
+    "wide-284",
+    "wide-285",
+    "wide-286",
+    "wide-287",
+    "wide-288",
+    "wide-289",
+    "wide-290",
+    "wide-291",
+    "wide-292",
+    "wide-293",
+    "wide-294",
+    "wide-295",
+    "wide-296",
+    "wide-297",
+    "wide-298",
+    "wide-299",
+    "wide-300",
+    "wide-301",
+    "wide-302",
+    "wide-303",
+    "wide-304",
+    "wide-305",
+    "wide-306",
+    "wide-307",
+    "wide-308",
+    "wide-309",
+    "wide-310",
+    "wide-311",
+    "wide-312",
+    "wide-313",
+    "wide-314",
+    "wide-315",
+    "wide-316",
+    "wide-317",
+    "wide-318",
+    "wide-319",
+    "wide-320"
+  ],
+  "summary": {
+    "questions": 320,
+    "source_checked_count": 320,
+    "all_facts_supported_count": 320,
+    "strict_count": 320,
+    "strict_accuracy": 1.0,
+    "error_count": 0,
+    "value_score": 1.0,
+    "entity_score": 1.0,
+    "relation_score": 1.0,
+    "property_score": 1.0,
+    "relation_choice_score": 1.0,
+    "document_recall": 1.0,
+    "source_recall": 1.0,
+    "binding_score": 1.0,
+    "path_score": 1.0,
+    "answerability_accuracy": 1.0,
+    "latency_mean_seconds": 0.081877,
+    "latency_median_seconds": 0.078039,
+    "latency_p95_seconds": 0.119541
+  }
+}

tools/graphrag/evaluation/results/final_v7_2026-10-26/summary.json

@@ -0,0 +1,270 @@
+{
+  "release": "graphrag-v7-frozen",
+  "graph_dev": {
+    "questions": 1046,
+    "source_checked_count": 1046,
+    "all_facts_supported_count": 1046,
+    "strict_count": 1046,
+    "strict_accuracy": 1.0,
+    "error_count": 0,
+    "value_score": 1.0,
+    "entity_score": 1.0,
+    "relation_score": 1.0,
+    "property_score": 1.0,
+    "relation_choice_score": 1.0,
+    "document_recall": 1.0,
+    "source_recall": 1.0,
+    "binding_score": 1.0,
+    "path_score": 1.0,
+    "answerability_accuracy": 1.0,
+    "latency_mean_seconds": 0.081848,
+    "latency_median_seconds": 0.071172,
+    "latency_p95_seconds": 0.113998
+  },
+  "graph_new_test": {
+    "questions": 80,
+    "source_checked_count": 80,
+    "all_facts_supported_count": 80,
+    "strict_count": 80,
+    "strict_accuracy": 1.0,
+    "error_count": 0,
+    "value_score": 1.0,
+    "entity_score": 1.0,
+    "relation_score": 1.0,
+    "property_score": 1.0,
+    "relation_choice_score": 1.0,
+    "document_recall": 1.0,
+    "source_recall": 1.0,
+    "binding_score": 1.0,
+    "path_score": 1.0,
+    "answerability_accuracy": 1.0,
+    "latency_mean_seconds": 0.093012,
+    "latency_median_seconds": 0.091794,
+    "latency_p95_seconds": 0.127876
+  },
+  "graph_historical_regression": {
+    "questions": 320,
+    "source_checked_count": 320,
+    "all_facts_supported_count": 320,
+    "strict_count": 320,
+    "strict_accuracy": 1.0,
+    "error_count": 0,
+    "value_score": 1.0,
+    "entity_score": 1.0,
+    "relation_score": 1.0,
+    "property_score": 1.0,
+    "relation_choice_score": 1.0,
+    "document_recall": 1.0,
+    "source_recall": 1.0,
+    "binding_score": 1.0,
+    "path_score": 1.0,
+    "answerability_accuracy": 1.0,
+    "latency_mean_seconds": 0.081877,
+    "latency_median_seconds": 0.078039,
+    "latency_p95_seconds": 0.119541
+  },
+  "model_dev": {
+    "questions": 46,
+    "correct": 46,
+    "accuracy": 1.0,
+    "errors": 0,
+    "first_pass_correct": 45,
+    "format_repair_attempted": 1,
+    "corrective_request_count": 1,
+    "context_repair_attempted": 0,
+    "normalized_first_response": 0
+  },
+  "model_new_test": {
+    "questions": 80,
+    "correct": 80,
+    "accuracy": 1.0,
+    "errors": 0,
+    "first_pass_correct": 77,
+    "format_repair_attempted": 1,
+    "corrective_request_count": 3,
+    "context_repair_attempted": 2,
+    "normalized_first_response": 0
+  },
+  "model_historical_regression": {
+    "questions": 320,
+    "correct": 319,
+    "accuracy": 0.996875,
+    "errors": 0,
+    "first_pass_correct": 310,
+    "format_repair_attempted": 3,
+    "corrective_request_count": 10,
+    "context_repair_attempted": 7,
+    "normalized_first_response": 0
+  },
+  "model_new_test_by_category": {
+    "candidate_set_aggregation": {
+      "questions": 7,
+      "correct": 7,
+      "accuracy": 1.0
+    },
+    "conjunctive_disambiguation": {
+      "questions": 6,
+      "correct": 6,
+      "accuracy": 1.0
+    },
+    "document_author_category_multi_hop": {
+      "questions": 2,
+      "correct": 2,
+      "accuracy": 1.0
+    },
+    "negative_lookup": {
+      "questions": 8,
+      "correct": 8,
+      "accuracy": 1.0
+    },
+    "person_keyword_title": {
+      "questions": 10,
+      "correct": 10,
+      "accuracy": 1.0
+    },
+    "person_multiple_categories": {
+      "questions": 7,
+      "correct": 7,
+      "accuracy": 1.0
+    },
+    "person_title_year": {
+      "questions": 10,
+      "correct": 10,
+      "accuracy": 1.0
+    },
+    "same_academic_year_pair": {
+      "questions": 10,
+      "correct": 10,
+      "accuracy": 1.0
+    },
+    "same_start_year_pair": {
+      "questions": 10,
+      "correct": 10,
+      "accuracy": 1.0
+    },
+    "shared_keyword_pair": {
+      "questions": 10,
+      "correct": 10,
+      "accuracy": 1.0
+    }
+  },
+  "model_regression_by_category": {
+    "candidate_set_aggregation": {
+      "questions": 32,
+      "correct": 32,
+      "accuracy": 1.0
+    },
+    "conjunctive_disambiguation": {
+      "questions": 40,
+      "correct": 39,
+      "accuracy": 0.975
+    },
+    "document_author_category_multi_hop": {
+      "questions": 24,
+      "correct": 24,
+      "accuracy": 1.0
+    },
+    "negative_lookup": {
+      "questions": 32,
+      "correct": 32,
+      "accuracy": 1.0
+    },
+    "person_keyword_title": {
+      "questions": 32,
+      "correct": 32,
+      "accuracy": 1.0
+    },
+    "person_multiple_categories": {
+      "questions": 24,
+      "correct": 24,
+      "accuracy": 1.0
+    },
+    "person_title_year": {
+      "questions": 24,
+      "correct": 24,
+      "accuracy": 1.0
+    },
+    "same_academic_year_pair": {
+      "questions": 40,
+      "correct": 40,
+      "accuracy": 1.0
+    },
+    "same_start_year_pair": {
+      "questions": 32,
+      "correct": 32,
+      "accuracy": 1.0
+    },
+    "shared_keyword_pair": {
+      "questions": 40,
+      "correct": 40,
+      "accuracy": 1.0
+    }
+  },
+  "protocol": {
+    "seed": 20261026,
+    "dev_count": 1046,
+    "fresh_test_count": 80,
+    "test_categories": {
+      "shared_keyword_pair": 10,
+      "same_academic_year_pair": 10,
+      "same_start_year_pair": 10,
+      "person_title_year": 10,
+      "person_keyword_title": 10,
+      "document_author_category_multi_hop": 2,
+      "person_multiple_categories": 7,
+      "candidate_set_aggregation": 7,
+      "conjunctive_disambiguation": 6,
+      "negative_lookup": 8
+    },
+    "test_positives": 72,
+    "test_negatives": 8,
+    "available_groups": {
+      "shared_keyword_pair": 1883,
+      "same_academic_year_pair": 535,
+      "same_start_year_pair": 153,
+      "person_title_year": 30,
+      "person_keyword_title": 146,
+      "document_author_category_multi_hop": 2,
+      "person_multiple_categories": 7,
+      "candidate_set_aggregation": 54,
+      "conjunctive_disambiguation": 628,
+      "negative_lookup": 106
+    },
+    "old_gold_corrections": 10,
+    "regression_count": 320,
+    "llm_dev_count": 46,
+    "target": 0.92,
+    "policy": "Freeze after complete DEV; run fresh TEST once, retain every failure. Regression remains separate.",
+    "annotation": "Independent raw-source oracle plus AI source review; not human review."
+  },
+  "tests": {
+    "passed": 591,
+    "skipped": 0
+  },
+  "source_review": {
+    "sample": 20,
+    "supported": 20,
+    "reviewer": "AI direct raw Markdown review, not human"
+  },
+  "functional_scenarios": {
+    "passed": 9,
+    "total": 9
+  },
+  "historical_rag_replay": {
+    "passed": 5,
+    "total": 5,
+    "new_generation": false
+  },
+  "failed_test_ids": [],
+  "failed_regression_ids": [
+    "wide-285"
+  ],
+  "limitations": [
+    "Known corpus, persons and template families; new query groups, not unseen profiles.",
+    "Source oracle plus AI review; no human annotation.",
+    "Only two fresh document-author/category groups and seven fresh category-list groups remain.",
+    "End-to-end native OpenWebUI tool routing is point 4.",
+    "Final prose is deterministic rendering of verified model-selected claims; this is not free-prose generation accuracy.",
+    "Gold corrections and changed delivery protocol prevent a controlled accuracy comparison with v6."
+  ]
+}

tools/graphrag/evaluation/results/final_v7_2026-10-26/verification.json

@@ -0,0 +1,42 @@
+{
+  "test_command": "RUN_NEO4J_TESTS=1 RUN_LIVE_TESTS=1 RUN_LIVE_RAG_E2E=1 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 python -m pytest -q",
+  "logs": [
+    {
+      "name": "zp-v7-release-tests.log",
+      "sha256": "17f87f311f897da7a4170b102894668f13f7078eec8213a33f501ac8ca012081",
+      "output": "........................................................................ [ 12%]\n........................................................................ [ 24%]\n........................................................................ [ 36%]\n........................................................................ [ 48%]\n........................................................................ [ 60%]\n........................................................................ [ 73%]\n........................................................................ [ 85%]\n........................................................................ [ 97%]\n...............                                                          [100%]\n=============================== warnings summary ===============================\n.venv/lib/python3.10/site-packages/fastapi/testclient.py:1\n  /home/janko/DP/zp-agent/.venv/lib/python3.10/site-packages/fastapi/testclient.py:1: StarletteDeprecationWarning: Using `httpx` with `starlette.testclient` is deprecated; install `httpx2` instead.\n    from starlette.testclient import TestClient as TestClient  # noqa\n\n-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html\n591 passed, 1 warning in 76.37s (0:01:16)\n"
+    },
+    {
+      "name": "zp-v7-targeted-final.log",
+      "sha256": "c53fc4867b28f28a687dcd19fbb04fb11dc7cc7c0a4a86abd2205b8233ad2b1c",
+      "output": "......................................................                   [100%]\n54 passed in 2.06s\n"
+    }
+  ],
+  "sync_response": {
+    "status": "ok",
+    "pull_git": false,
+    "duration_seconds": 163.9,
+    "counts": {
+      "documents": 114,
+      "chunks": 887,
+      "fts_chunks": 887,
+      "embedding_chunks": 887,
+      "tags": 2189,
+      "categories": 1193
+    },
+    "index_version": {
+      "version": "33ea0794472c416e97c84371a2a6cdf8",
+      "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+      "created_at_utc": "2026-10-26T00:10:27.444513+00:00",
+      "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+      "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+      "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+      "embedding_model": "intfloat/multilingual-e5-small",
+      "publication_policy": "explicit-false-hidden-v1"
+    },
+    "graph_counts": {
+      "nodes": 446,
+      "relationships": 877
+    }
+  }
+}

tools/graphrag/evaluation/results/v7_preparation/generation_dev_ids.json

@@ -0,0 +1,48 @@
+[
+  "v7-dev-year-1899",
+  "v7-dev-year-1999",
+  "v7-dev-year-2100",
+  "v7-dev-year-9999",
+  "wide-001",
+  "wide-002",
+  "wide-041",
+  "wide-042",
+  "wide-080",
+  "wide-081",
+  "wide-082",
+  "wide-104",
+  "wide-113",
+  "wide-114",
+  "wide-116",
+  "wide-119",
+  "wide-129",
+  "wide-137",
+  "wide-138",
+  "wide-139",
+  "wide-141",
+  "wide-142",
+  "wide-143",
+  "wide-145",
+  "wide-155",
+  "wide-159",
+  "wide-161",
+  "wide-163",
+  "wide-165",
+  "wide-169",
+  "wide-170",
+  "wide-178",
+  "wide-181",
+  "wide-193",
+  "wide-194",
+  "wide-209",
+  "wide-217",
+  "wide-218",
+  "wide-232",
+  "wide-236",
+  "wide-239",
+  "wide-249",
+  "wide-250",
+  "wide-283",
+  "wide-289",
+  "wide-290"
+]

tools/graphrag/evaluation/results/v7_preparation/live_scenarios.declaration.json

@@ -0,0 +1,417 @@
+{
+  "method": "Live synthetic functional checks; separate from corpus accuracy metrics.",
+  "model": "model120-fast",
+  "hashes": {
+    "/home/janko/DP/zp-agent/tools/answer_evaluation/check_context_scenarios.py": "99dbae2747183622e0823cd8f6536f4ff38ef56d0e41dd7855096d47af3b1c37",
+    "/home/janko/DP/zp-agent/tools/answer_evaluation/graph_answer_prompt.txt": "fc9f2f8d52e70a8d67b9f4814b3f15e6be64bf88afe53ce4db4f037bde00e064",
+    "/home/janko/DP/zp-agent/tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47",
+    "/home/janko/DP/zp-agent/tools/answer_evaluation/checked_generation.py": "cb821a65327424c19cbd996190b14e6aa30d9b909493f11716552f71249a0529",
+    "/home/janko/DP/zp-agent/tools/answer_evaluation/answer_format.py": "86a27b4aebc6a5077e13794b81957a1cf063a1f3e070c583120ba53c1788d2f8",
+    "/home/janko/DP/zp-agent/tools/answer_evaluation/render_answer.py": "609c534a37899b3febcea32d983da4b806c287de8f2ddf9bc0675e3280859002"
+  },
+  "scenarios": [
+    {
+      "id": "complete",
+      "question": "Uveď názov práce osoby Eva Vzorová a rok práce.",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": true,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026
+          }
+        ]
+      }
+    },
+    {
+      "id": "incomplete_without_truncated",
+      "question": "Uveď názov práce osoby Eva Vzorová a rok práce.",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": false,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026
+          }
+        ]
+      }
+    },
+    {
+      "id": "truncated",
+      "question": "Uveď názov práce osoby Eva Vzorová a rok práce.",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": false,
+        "truncated": true,
+        "truncation_reasons": [
+          "limit"
+        ],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026
+          }
+        ]
+      }
+    },
+    {
+      "id": "unknown_completeness",
+      "question": "Uveď názov práce osoby Eva Vzorová a rok práce.",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": null,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026
+          }
+        ]
+      }
+    },
+    {
+      "id": "exact_count_partial_evidence",
+      "question": "Aký je počet vyhovujúcich osôb?",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": false,
+        "truncated": true,
+        "truncation_reasons": [
+          "evidence_limit"
+        ],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "aggregation": {
+          "count": 2,
+          "complete": true
+        },
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "ok",
+        "count": 2
+      }
+    },
+    {
+      "id": "same_person_child_tag",
+      "question": "Uveď názov práce osoby Eva Vzorová v roku 2026 s kľúčovým slovom python.",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": true,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          },
+          {
+            "source_id": "S2",
+            "document_path": "child.md",
+            "source_url": "https://example.test/child"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          },
+          {
+            "person": "Eva Vzorová",
+            "keyword": "python",
+            "document_path": "child.md",
+            "source_url": "https://example.test/child"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "keyword": "python",
+            "evidence_documents": [
+              "a.md",
+              "child.md"
+            ],
+            "source_urls": [
+              "https://example.test/a",
+              "https://example.test/child"
+            ]
+          }
+        ]
+      },
+      "expected": {
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "keyword": "python"
+          }
+        ]
+      }
+    },
+    {
+      "id": "same_person_year_child_tag",
+      "question": "Ktorá osoba spĺňa kľúčové slovo python a rok práce 2026?",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": true,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          },
+          {
+            "source_id": "S2",
+            "document_path": "child.md",
+            "source_url": "https://example.test/child"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          },
+          {
+            "person": "Eva Vzorová",
+            "keyword": "python",
+            "document_path": "child.md",
+            "source_url": "https://example.test/child"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": [
+          {
+            "person": "Eva Vzorová",
+            "year": 2026,
+            "keyword": "python",
+            "evidence_documents": [
+              "a.md",
+              "child.md"
+            ],
+            "source_urls": [
+              "https://example.test/a",
+              "https://example.test/child"
+            ]
+          }
+        ]
+      },
+      "expected": {
+        "status": "ok",
+        "claims": [
+          {
+            "person": "Eva Vzorová",
+            "year": 2026,
+            "keyword": "python"
+          }
+        ]
+      }
+    },
+    {
+      "id": "type_is_not_keyword",
+      "question": "Uveď názov diplomového projektu osoby Eva Vzorová a rok.",
+      "context": {
+        "status": "ok",
+        "reason": null,
+        "complete": true,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [
+          {
+            "source_id": "S1",
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "facts": [
+          {
+            "person": "Eva Vzorová",
+            "title": "Vyhľadávanie informácií",
+            "year": 2026,
+            "document_path": "a.md",
+            "source_url": "https://example.test/a"
+          }
+        ],
+        "aggregation": null,
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "ok",
+        "no_keyword": true
+      }
+    },
+    {
+      "id": "no_answer",
+      "question": "Akú prácu má Neexistujúca Osoba?",
+      "context": {
+        "status": "no_answer",
+        "reason": "unknown_person",
+        "complete": true,
+        "truncated": false,
+        "truncation_reasons": [],
+        "index_version": {
+          "version": "synthetic-functional-fixture"
+        },
+        "sources": [],
+        "facts": [],
+        "aggregation": null,
+        "answer_candidates": []
+      },
+      "expected": {
+        "status": "no_answer"
+      }
+    }
+  ]
+}

tools/graphrag/evaluation/results/v7_preparation/protocol.json

@@ -0,0 +1,37 @@
+{
+  "seed": 20261026,
+  "dev_count": 1046,
+  "fresh_test_count": 80,
+  "test_categories": {
+    "shared_keyword_pair": 10,
+    "same_academic_year_pair": 10,
+    "same_start_year_pair": 10,
+    "person_title_year": 10,
+    "person_keyword_title": 10,
+    "document_author_category_multi_hop": 2,
+    "person_multiple_categories": 7,
+    "candidate_set_aggregation": 7,
+    "conjunctive_disambiguation": 6,
+    "negative_lookup": 8
+  },
+  "test_positives": 72,
+  "test_negatives": 8,
+  "available_groups": {
+    "shared_keyword_pair": 1883,
+    "same_academic_year_pair": 535,
+    "same_start_year_pair": 153,
+    "person_title_year": 30,
+    "person_keyword_title": 146,
+    "document_author_category_multi_hop": 2,
+    "person_multiple_categories": 7,
+    "candidate_set_aggregation": 54,
+    "conjunctive_disambiguation": 628,
+    "negative_lookup": 106
+  },
+  "old_gold_corrections": 10,
+  "regression_count": 320,
+  "llm_dev_count": 46,
+  "target": 0.92,
+  "policy": "Freeze after complete DEV; run fresh TEST once, retain every failure. Regression remains separate.",
+  "annotation": "Independent raw-source oracle plus AI source review; not human review."
+}

tools/graphrag/evaluation/results/v7_preparation/rag_delivery_replay.json

@@ -0,0 +1,478 @@
+{
+  "configuration": {
+    "method": "Deterministic delivery replay on the same saved model responses, not new generation.",
+    "input": "tools/graphrag/evaluation/results/v6_preparation/rag_live_smoke.json",
+    "input_sha256": "674272fd911a5ae04690150ac899531feae90b4d44a1735a530a2838a0d32ff3",
+    "delivery_files": {
+      "/home/janko/DP/zp-agent/tools/answer_evaluation/rag_delivery.py": "b83f93d446f44fdd7d41c12eeeaae69c0b02908076624f4eb3f4f72b9b88dbc8",
+      "tools/rag/evaluation/evaluate_rag_answers.py": "945c2da655537fb37fe82a08e10dc3ed66a2d90980921f61f61efb171a4a26a2",
+      "tools/rag/evaluation/rag_metrics.py": "3e95922df90dd1b757a550040a2d9bd4f857e17d179f60170d7cc881a08f94b6",
+      "tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1",
+      "tools/answer_evaluation/context_policy.py": "0dc2a056ca0d886d7cc8248ed4632c910c5e35bebd03ef94573df41802297f47"
+    }
+  },
+  "results": [
+    {
+      "id": "q0053",
+      "question": "Aký je názov diplomová práca osoby Lukáš Pokrývka?",
+      "citations_ok": true,
+      "unsupported_urls": [],
+      "retrieval_coverage": [
+        {
+          "complete": false,
+          "truncated": true,
+          "truncation_reasons": [
+            "candidate_limit",
+            "per_document_limit",
+            "result_limit"
+          ],
+          "index_version": {
+            "version": "10b78a8e70744e2792ffe4cf11d45571",
+            "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+            "created_at_utc": "2026-10-25T21:10:39.160542+00:00",
+            "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+            "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+            "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+            "embedding_model": "intfloat/multilingual-e5-small",
+            "publication_policy": "explicit-false-hidden-v1"
+          }
+        }
+      ],
+      "context_issues": [],
+      "accepted_for_delivery": true,
+      "answer": "Výpis je neúplný.\n\nNázov diplomovej práce Lukáša Pokrývku je „Paralelné trénovanie neurónových sietí\".\n\nZdroj: https://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+      "tool_called": true,
+      "tool_call_count": 1,
+      "tool_calls": [
+        {
+          "name": "retrieve_zpwiki_context",
+          "arguments": {
+            "query": "Aký je názov diplomová práca osoby Lukáš Pokrývka?",
+            "limit": 5,
+            "published_only": true,
+            "max_per_document": 1
+          },
+          "latency_seconds": 8.6679,
+          "source_urls": [
+            "https://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+            "https://zp.kemt.fei.tuke.sk/students/2016/tomas_kucharik",
+            "https://zp.kemt.fei.tuke.sk/students/2019/patrik_pokrivcak",
+            "https://zp.kemt.fei.tuke.sk/students/2018/oleh_bilykh",
+            "https://zp.kemt.fei.tuke.sk/students/2016/jakub_maruniak/dp2021"
+          ]
+        }
+      ],
+      "rag_source_urls": [
+        "https://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+        "https://zp.kemt.fei.tuke.sk/students/2016/tomas_kucharik",
+        "https://zp.kemt.fei.tuke.sk/students/2019/patrik_pokrivcak",
+        "https://zp.kemt.fei.tuke.sk/students/2018/oleh_bilykh",
+        "https://zp.kemt.fei.tuke.sk/students/2016/jakub_maruniak/dp2021"
+      ],
+      "first_model_latency_seconds": 5.65362,
+      "tool_latency_seconds": 8.6679,
+      "final_model_latency_seconds": 23.053733,
+      "total_latency_seconds": 37.375855,
+      "usage": {
+        "prompt_tokens": 19045,
+        "completion_tokens": 137,
+        "total_tokens": 19182
+      },
+      "response_model": "Qwen3.5-122B-A10B-Q4_K_M",
+      "citations_checked": true,
+      "answer_matches": [
+        true
+      ],
+      "answer_contains_score": 1.0,
+      "source_matches": [
+        true
+      ],
+      "source_match_count": 1,
+      "source_required_count": 1,
+      "source_url_score": 1.0,
+      "should_answer_ok": true,
+      "should_answer_score": 1.0,
+      "returned_no_answer": false,
+      "tool_score": 1.0,
+      "overall_score": 1.0,
+      "strict_pass": true,
+      "model_answer": "Názov diplomovej práce Lukáša Pokrývku je „Paralelné trénovanie neurónových sietí\".\n\nZdroj: https://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+      "first_pass_accepted_for_delivery": false,
+      "first_pass_context_issues": [
+        "missing_incompleteness_notice"
+      ],
+      "display_notice_added": true
+    },
+    {
+      "id": "q0060",
+      "question": "Rozlíš rok začiatku štúdia a rok práce osoby Lukáš Pokrývka.",
+      "citations_ok": true,
+      "unsupported_urls": [],
+      "retrieval_coverage": [
+        {
+          "complete": false,
+          "truncated": true,
+          "truncation_reasons": [
+            "candidate_limit",
+            "per_document_limit",
+            "result_limit"
+          ],
+          "index_version": {
+            "version": "10b78a8e70744e2792ffe4cf11d45571",
+            "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+            "created_at_utc": "2026-10-25T21:10:39.160542+00:00",
+            "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+            "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+            "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+            "embedding_model": "intfloat/multilingual-e5-small",
+            "publication_policy": "explicit-false-hidden-v1"
+          }
+        }
+      ],
+      "context_issues": [],
+      "accepted_for_delivery": true,
+      "answer": "Výpis je neúplný.\n\nRok začiatku štúdia osoby Lukáš Pokrývka je 2016, zatiaľ čo jeho diplomová práca bola realizovaná v roku 2021.\n\nZdroj:\nhttps://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+      "tool_called": true,
+      "tool_call_count": 1,
+      "tool_calls": [
+        {
+          "name": "retrieve_zpwiki_context",
+          "arguments": {
+            "query": "Rozlíš rok začiatku štúdia a rok práce osoby Lukáš Pokrývka.",
+            "limit": 10,
+            "published_only": true,
+            "max_per_document": 5
+          },
+          "latency_seconds": 8.727633,
+          "source_urls": [
+            "https://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+            "https://zp.kemt.fei.tuke.sk/students/2020/matej_kobylan",
+            "https://zp.kemt.fei.tuke.sk/students/2020/jakub_kristian_lukas",
+            "https://zp.kemt.fei.tuke.sk/students/2019/michal_stromko",
+            "https://zp.kemt.fei.tuke.sk/students/2018/ondrej_megela"
+          ]
+        }
+      ],
+      "rag_source_urls": [
+        "https://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+        "https://zp.kemt.fei.tuke.sk/students/2020/matej_kobylan",
+        "https://zp.kemt.fei.tuke.sk/students/2020/jakub_kristian_lukas",
+        "https://zp.kemt.fei.tuke.sk/students/2019/michal_stromko",
+        "https://zp.kemt.fei.tuke.sk/students/2018/ondrej_megela"
+      ],
+      "first_model_latency_seconds": 5.710692,
+      "tool_latency_seconds": 8.727633,
+      "final_model_latency_seconds": 40.777827,
+      "total_latency_seconds": 55.216681,
+      "usage": {
+        "prompt_tokens": 20160,
+        "completion_tokens": 155,
+        "total_tokens": 20315
+      },
+      "response_model": "Qwen3.5-122B-A10B-Q4_K_M",
+      "citations_checked": true,
+      "answer_matches": [
+        true,
+        true
+      ],
+      "answer_contains_score": 1.0,
+      "source_matches": [
+        true
+      ],
+      "source_match_count": 1,
+      "source_required_count": 1,
+      "source_url_score": 1.0,
+      "should_answer_ok": true,
+      "should_answer_score": 1.0,
+      "returned_no_answer": false,
+      "tool_score": 1.0,
+      "overall_score": 1.0,
+      "strict_pass": true,
+      "model_answer": "Rok začiatku štúdia osoby Lukáš Pokrývka je 2016, zatiaľ čo jeho diplomová práca bola realizovaná v roku 2021.\n\nZdroj:\nhttps://zp.kemt.fei.tuke.sk/students/2016/lukas_pokryvka",
+      "first_pass_accepted_for_delivery": false,
+      "first_pass_context_issues": [
+        "missing_incompleteness_notice"
+      ],
+      "display_notice_added": true
+    },
+    {
+      "id": "q0608",
+      "question": "Aký source_url patrí k stránke „Podpora slovenčiny v knižnici Spacy“?",
+      "citations_ok": true,
+      "unsupported_urls": [],
+      "retrieval_coverage": [
+        {
+          "complete": false,
+          "truncated": true,
+          "truncation_reasons": [
+            "candidate_limit",
+            "per_document_limit",
+            "result_limit"
+          ],
+          "index_version": {
+            "version": "10b78a8e70744e2792ffe4cf11d45571",
+            "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+            "created_at_utc": "2026-10-25T21:10:39.160542+00:00",
+            "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+            "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+            "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+            "embedding_model": "intfloat/multilingual-e5-small",
+            "publication_policy": "explicit-false-hidden-v1"
+          }
+        }
+      ],
+      "context_issues": [],
+      "accepted_for_delivery": true,
+      "answer": "Výpis je neúplný.\n\nStránka s názvom „Podpora slovenčiny v knižnici Spacy\" má pridelený zdrojový URL adresu https://zp.kemt.fei.tuke.sk/topics/spacy.\n\nZdroj: https://zp.kemt.fei.tuke.sk/topics/spacy",
+      "tool_called": true,
+      "tool_call_count": 1,
+      "tool_calls": [
+        {
+          "name": "retrieve_zpwiki_context",
+          "arguments": {
+            "query": "Aký source_url patrí k stránke „Podpora slovenčiny v knižnici Spacy“?",
+            "limit": 5,
+            "max_per_document": 1,
+            "published_only": true
+          },
+          "latency_seconds": 0.620349,
+          "source_urls": [
+            "https://zp.kemt.fei.tuke.sk/topics/spacy",
+            "https://zp.kemt.fei.tuke.sk/students/2016/maros_harahus",
+            "https://zp.kemt.fei.tuke.sk/students/2016/jakub_maruniak",
+            "https://zp.kemt.fei.tuke.sk/students/2017/than_trung_thanh",
+            "https://zp.kemt.fei.tuke.sk/topics/resources"
+          ]
+        }
+      ],
+      "rag_source_urls": [
+        "https://zp.kemt.fei.tuke.sk/topics/spacy",
+        "https://zp.kemt.fei.tuke.sk/students/2016/maros_harahus",
+        "https://zp.kemt.fei.tuke.sk/students/2016/jakub_maruniak",
+        "https://zp.kemt.fei.tuke.sk/students/2017/than_trung_thanh",
+        "https://zp.kemt.fei.tuke.sk/topics/resources"
+      ],
+      "first_model_latency_seconds": 28.658617,
+      "tool_latency_seconds": 0.620349,
+      "final_model_latency_seconds": 35.815477,
+      "total_latency_seconds": 65.094973,
+      "usage": {
+        "prompt_tokens": 28510,
+        "completion_tokens": 141,
+        "total_tokens": 28651
+      },
+      "response_model": "Qwen3.5-122B-A10B-Q4_K_M",
+      "citations_checked": true,
+      "answer_matches": [
+        true
+      ],
+      "answer_contains_score": 1.0,
+      "source_matches": [
+        true
+      ],
+      "source_match_count": 1,
+      "source_required_count": 1,
+      "source_url_score": 1.0,
+      "should_answer_ok": true,
+      "should_answer_score": 1.0,
+      "returned_no_answer": false,
+      "tool_score": 1.0,
+      "overall_score": 1.0,
+      "strict_pass": true,
+      "model_answer": "Stránka s názvom „Podpora slovenčiny v knižnici Spacy\" má pridelený zdrojový URL adresu https://zp.kemt.fei.tuke.sk/topics/spacy.\n\nZdroj: https://zp.kemt.fei.tuke.sk/topics/spacy",
+      "first_pass_accepted_for_delivery": false,
+      "first_pass_context_issues": [
+        "missing_incompleteness_notice"
+      ],
+      "display_notice_added": true
+    },
+    {
+      "id": "q1147",
+      "question": "Porovnaj záznamy osôb Ján Holp a Ondrej Megela. Pri osobe Ján Holp použi záznam typu „diplomová práca“ a pri osobe Ondrej Megela záznam typu „bakalárska práca“. Uveď názov každej práce a rok, pri ktorom je na ZP Wiki uvedená.",
+      "citations_ok": true,
+      "unsupported_urls": [],
+      "retrieval_coverage": [
+        {
+          "complete": false,
+          "truncated": true,
+          "truncation_reasons": [
+            "candidate_limit",
+            "per_document_limit",
+            "result_limit"
+          ],
+          "index_version": {
+            "version": "10b78a8e70744e2792ffe4cf11d45571",
+            "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+            "created_at_utc": "2026-10-25T21:10:39.160542+00:00",
+            "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+            "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+            "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+            "embedding_model": "intfloat/multilingual-e5-small",
+            "publication_policy": "explicit-false-hidden-v1"
+          }
+        }
+      ],
+      "context_issues": [],
+      "accepted_for_delivery": true,
+      "answer": "Výpis je neúplný.\n\nJán Holp vypracoval diplomovú prácu s názvom „Systém získavania informácií v slovenskom jazyku\" v roku 2021 a Ondrej Megela vypracoval bakalársku prácu s názvom „Neurónové jazykové modelovanie typu BERT\" v roku 2020.\n\nZdroje:\nhttps://zp.kemt.fei.tuke.sk/students/2016/jan_holp\nhttps://zp.kemt.fei.tuke.sk/students/2018/ondrej_megela",
+      "tool_called": true,
+      "tool_call_count": 1,
+      "tool_calls": [
+        {
+          "name": "retrieve_zpwiki_context",
+          "arguments": {
+            "query": "Porovnaj záznamy osôb Ján Holp a Ondrej Megela. Pri osobe Ján Holp použi záznam typu „diplomová práca“ a pri osobe Ondrej Megela záznam typu „bakalárska práca“. Uveď názov každej práce a rok, pri ktorom je na ZP Wiki uvedená.",
+            "limit": 5,
+            "published_only": true,
+            "max_per_document": 1
+          },
+          "latency_seconds": 1.001685,
+          "source_urls": [
+            "https://zp.kemt.fei.tuke.sk/students/2018/ondrej_megela",
+            "https://zp.kemt.fei.tuke.sk/students/2016/jan_holp",
+            "https://zp.kemt.fei.tuke.sk/students/2019/michal_stromko",
+            "https://zp.kemt.fei.tuke.sk/topics/faq",
+            "https://zp.kemt.fei.tuke.sk/topics/bert"
+          ]
+        }
+      ],
+      "rag_source_urls": [
+        "https://zp.kemt.fei.tuke.sk/students/2018/ondrej_megela",
+        "https://zp.kemt.fei.tuke.sk/students/2016/jan_holp",
+        "https://zp.kemt.fei.tuke.sk/students/2019/michal_stromko",
+        "https://zp.kemt.fei.tuke.sk/topics/faq",
+        "https://zp.kemt.fei.tuke.sk/topics/bert"
+      ],
+      "first_model_latency_seconds": 14.535465,
+      "tool_latency_seconds": 1.001685,
+      "final_model_latency_seconds": 56.647723,
+      "total_latency_seconds": 72.185893,
+      "usage": {
+        "prompt_tokens": 25893,
+        "completion_tokens": 208,
+        "total_tokens": 26101
+      },
+      "response_model": "Qwen3.5-122B-A10B-Q4_K_M",
+      "citations_checked": true,
+      "answer_matches": [
+        true,
+        true,
+        true,
+        true
+      ],
+      "answer_contains_score": 1.0,
+      "source_matches": [
+        true,
+        true
+      ],
+      "source_match_count": 2,
+      "source_required_count": 2,
+      "source_url_score": 1.0,
+      "should_answer_ok": true,
+      "should_answer_score": 1.0,
+      "returned_no_answer": false,
+      "tool_score": 1.0,
+      "overall_score": 1.0,
+      "strict_pass": true,
+      "model_answer": "Ján Holp vypracoval diplomovú prácu s názvom „Systém získavania informácií v slovenskom jazyku\" v roku 2021 a Ondrej Megela vypracoval bakalársku prácu s názvom „Neurónové jazykové modelovanie typu BERT\" v roku 2020.\n\nZdroje:\nhttps://zp.kemt.fei.tuke.sk/students/2016/jan_holp\nhttps://zp.kemt.fei.tuke.sk/students/2018/ondrej_megela",
+      "first_pass_accepted_for_delivery": false,
+      "first_pass_context_issues": [
+        "missing_incompleteness_notice"
+      ],
+      "display_notice_added": true
+    },
+    {
+      "id": "q0888",
+      "question": "Aký konkrétny finančný rozpočet bol uvedený pre projekt osoby Tomáš Kuchárik?",
+      "citations_ok": true,
+      "unsupported_urls": [],
+      "retrieval_coverage": [
+        {
+          "complete": false,
+          "truncated": true,
+          "truncation_reasons": [
+            "candidate_limit",
+            "per_document_limit",
+            "result_limit"
+          ],
+          "index_version": {
+            "version": "10b78a8e70744e2792ffe4cf11d45571",
+            "git_revision": "3e892a1d49e9bce118cd03478ce15e01c2bebb4e",
+            "created_at_utc": "2026-10-25T21:10:39.160542+00:00",
+            "source_sha256": "007792c12f3467dd3203bb4951ccc197fdc0eef669580f1b1f893baf609ca363",
+            "documents_sha256": "0542af2a373771cf7c4a95f93d14f4f687f6a0187846892c959aa6efc4bade52",
+            "chunks_sha256": "437dde589be46e0c30842dd330de22d96a0173d2dbaff9c5f129568cb2bca4c9",
+            "embedding_model": "intfloat/multilingual-e5-small",
+            "publication_policy": "explicit-false-hidden-v1"
+          }
+        }
+      ],
+      "context_issues": [],
+      "accepted_for_delivery": true,
+      "answer": "Výpis je neúplný.\n\nV dostupných dokumentoch ZP Wiki sa informácia o konkrétnom finančnom rozpočte pre projekt osoby Tomáš Kuchárik nepodarilo spoľahlivo nájsť.\n\nZdroj:\nhttps://zp.kemt.fei.tuke.sk/students/2016/tomas_kucharik",
+      "tool_called": true,
+      "tool_call_count": 1,
+      "tool_calls": [
+        {
+          "name": "retrieve_zpwiki_context",
+          "arguments": {
+            "query": "Aký konkrétny finančný rozpočet bol uvedený pre projekt osoby Tomáš Kuchárik?",
+            "limit": 5,
+            "published_only": true
+          },
+          "latency_seconds": 0.454629,
+          "source_urls": [
+            "https://zp.kemt.fei.tuke.sk/students/2016/tomas_kucharik",
+            "https://zp.kemt.fei.tuke.sk/students/2020/matej_kobylan",
+            "https://zp.kemt.fei.tuke.sk/students/2019/filip_tomas",
+            "https://zp.kemt.fei.tuke.sk/students/2020/matus_suchanic",
+            "https://zp.kemt.fei.tuke.sk/students/2019/michal_stromko"
+          ]
+        }
+      ],
+      "rag_source_urls": [
+        "https://zp.kemt.fei.tuke.sk/students/2016/tomas_kucharik",
+        "https://zp.kemt.fei.tuke.sk/students/2020/matej_kobylan",
+        "https://zp.kemt.fei.tuke.sk/students/2019/filip_tomas",
+        "https://zp.kemt.fei.tuke.sk/students/2020/matus_suchanic",
+        "https://zp.kemt.fei.tuke.sk/students/2019/michal_stromko"
+      ],
+      "first_model_latency_seconds": 129.609931,
+      "tool_latency_seconds": 0.454629,
+      "final_model_latency_seconds": 20.02263,
+      "total_latency_seconds": 150.087617,
+      "usage": {
+        "prompt_tokens": 14022,
+        "completion_tokens": 136,
+        "total_tokens": 14158
+      },
+      "response_model": "Qwen3.5-122B-A10B-Q4_K_M",
+      "citations_checked": true,
+      "answer_matches": [],
+      "answer_contains_score": 1.0,
+      "source_matches": [
+        true
+      ],
+      "source_match_count": 1,
+      "source_required_count": 0,
+      "source_url_score": 1.0,
+      "should_answer_ok": true,
+      "should_answer_score": 1.0,
+      "returned_no_answer": true,
+      "tool_score": 1.0,
+      "overall_score": 1.0,
+      "strict_pass": true,
+      "model_answer": "V dostupných dokumentoch ZP Wiki sa informácia o konkrétnom finančnom rozpočte pre projekt osoby Tomáš Kuchárik nepodarilo spoľahlivo nájsť.\n\nZdroj:\nhttps://zp.kemt.fei.tuke.sk/students/2016/tomas_kucharik",
+      "first_pass_accepted_for_delivery": false,
+      "first_pass_context_issues": [
+        "missing_incompleteness_notice"
+      ],
+      "display_notice_added": true
+    }
+  ],
+  "summary": {
+    "questions": 5,
+    "strict_pass": 5,
+    "display_notice_added": 5
+  }
+}

tools/graphrag/evaluation/results/v7_preparation/regression_ids.json

@@ -0,0 +1,322 @@
+[
+  "wide-001",
+  "wide-002",
+  "wide-003",
+  "wide-004",
+  "wide-005",
+  "wide-006",
+  "wide-007",
+  "wide-008",
+  "wide-009",
+  "wide-010",
+  "wide-011",
+  "wide-012",
+  "wide-013",
+  "wide-014",
+  "wide-015",
+  "wide-016",
+  "wide-017",
+  "wide-018",
+  "wide-019",
+  "wide-020",
+  "wide-021",
+  "wide-022",
+  "wide-023",
+  "wide-024",
+  "wide-025",
+  "wide-026",
+  "wide-027",
+  "wide-028",
+  "wide-029",
+  "wide-030",
+  "wide-031",
+  "wide-032",
+  "wide-033",
+  "wide-034",
+  "wide-035",
+  "wide-036",
+  "wide-037",
+  "wide-038",
+  "wide-039",
+  "wide-040",
+  "wide-041",
+  "wide-042",
+  "wide-043",
+  "wide-044",
+  "wide-045",
+  "wide-046",
+  "wide-047",
+  "wide-048",
+  "wide-049",
+  "wide-050",
+  "wide-051",
+  "wide-052",
+  "wide-053",
+  "wide-054",
+  "wide-055",
+  "wide-056",
+  "wide-057",
+  "wide-058",
+  "wide-059",
+  "wide-060",
+  "wide-061",
+  "wide-062",
+  "wide-063",
+  "wide-064",
+  "wide-065",
+  "wide-066",
+  "wide-067",
+  "wide-068",
+  "wide-069",
+  "wide-070",
+  "wide-071",
+  "wide-072",
+  "wide-073",
+  "wide-074",
+  "wide-075",
+  "wide-076",
+  "wide-077",
+  "wide-078",
+  "wide-079",
+  "wide-080",
+  "wide-081",
+  "wide-082",
+  "wide-083",
+  "wide-084",
+  "wide-085",
+  "wide-086",
+  "wide-087",
+  "wide-088",
+  "wide-089",
+  "wide-090",
+  "wide-091",
+  "wide-092",
+  "wide-093",
+  "wide-094",
+  "wide-095",
+  "wide-096",
+  "wide-097",
+  "wide-098",
+  "wide-099",
+  "wide-100",
+  "wide-101",
+  "wide-102",
+  "wide-103",
+  "wide-104",
+  "wide-105",
+  "wide-106",
+  "wide-107",
+  "wide-108",
+  "wide-109",
+  "wide-110",
+  "wide-111",
+  "wide-112",
+  "wide-113",
+  "wide-114",
+  "wide-115",
+  "wide-116",
+  "wide-117",
+  "wide-118",
+  "wide-119",
+  "wide-120",
+  "wide-121",
+  "wide-122",
+  "wide-123",
+  "wide-124",
+  "wide-125",
+  "wide-126",
+  "wide-127",
+  "wide-128",
+  "wide-129",
+  "wide-130",
+  "wide-131",
+  "wide-132",
+  "wide-133",
+  "wide-134",
+  "wide-135",
+  "wide-136",
+  "wide-137",
+  "wide-138",
+  "wide-139",
+  "wide-140",
+  "wide-141",
+  "wide-142",
+  "wide-143",
+  "wide-144",
+  "wide-145",
+  "wide-146",
+  "wide-147",
+  "wide-148",
+  "wide-149",
+  "wide-150",
+  "wide-151",
+  "wide-152",
+  "wide-153",
+  "wide-154",
+  "wide-155",
+  "wide-156",
+  "wide-157",
+  "wide-158",
+  "wide-159",
+  "wide-160",
+  "wide-161",
+  "wide-162",
+  "wide-163",
+  "wide-164",
+  "wide-165",
+  "wide-166",
+  "wide-167",
+  "wide-168",
+  "wide-169",
+  "wide-170",
+  "wide-171",
+  "wide-172",
+  "wide-173",
+  "wide-174",
+  "wide-175",
+  "wide-176",
+  "wide-177",
+  "wide-178",
+  "wide-179",
+  "wide-180",
+  "wide-181",
+  "wide-182",
+  "wide-183",
+  "wide-184",
+  "wide-185",
+  "wide-186",
+  "wide-187",
+  "wide-188",
+  "wide-189",
+  "wide-190",
+  "wide-191",
+  "wide-192",
+  "wide-193",
+  "wide-194",
+  "wide-195",
+  "wide-196",
+  "wide-197",
+  "wide-198",
+  "wide-199",
+  "wide-200",
+  "wide-201",
+  "wide-202",
+  "wide-203",
+  "wide-204",
+  "wide-205",
+  "wide-206",
+  "wide-207",
+  "wide-208",
+  "wide-209",
+  "wide-210",
+  "wide-211",
+  "wide-212",
+  "wide-213",
+  "wide-214",
+  "wide-215",
+  "wide-216",
+  "wide-217",
+  "wide-218",
+  "wide-219",
+  "wide-220",
+  "wide-221",
+  "wide-222",
+  "wide-223",
+  "wide-224",
+  "wide-225",
+  "wide-226",
+  "wide-227",
+  "wide-228",
+  "wide-229",
+  "wide-230",
+  "wide-231",
+  "wide-232",
+  "wide-233",
+  "wide-234",
+  "wide-235",
+  "wide-236",
+  "wide-237",
+  "wide-238",
+  "wide-239",
+  "wide-240",
+  "wide-241",
+  "wide-242",
+  "wide-243",
+  "wide-244",
+  "wide-245",
+  "wide-246",
+  "wide-247",
+  "wide-248",
+  "wide-249",
+  "wide-250",
+  "wide-251",
+  "wide-252",
+  "wide-253",
+  "wide-254",
+  "wide-255",
+  "wide-256",
+  "wide-257",
+  "wide-258",
+  "wide-259",
+  "wide-260",
+  "wide-261",
+  "wide-262",
+  "wide-263",
+  "wide-264",
+  "wide-265",
+  "wide-266",
+  "wide-267",
+  "wide-268",
+  "wide-269",
+  "wide-270",
+  "wide-271",
+  "wide-272",
+  "wide-273",
+  "wide-274",
+  "wide-275",
+  "wide-276",
+  "wide-277",
+  "wide-278",
+  "wide-279",
+  "wide-280",
+  "wide-281",
+  "wide-282",
+  "wide-283",
+  "wide-284",
+  "wide-285",
+  "wide-286",
+  "wide-287",
+  "wide-288",
+  "wide-289",
+  "wide-290",
+  "wide-291",
+  "wide-292",
+  "wide-293",
+  "wide-294",
+  "wide-295",
+  "wide-296",
+  "wide-297",
+  "wide-298",
+  "wide-299",
+  "wide-300",
+  "wide-301",
+  "wide-302",
+  "wide-303",
+  "wide-304",
+  "wide-305",
+  "wide-306",
+  "wide-307",
+  "wide-308",
+  "wide-309",
+  "wide-310",
+  "wide-311",
+  "wide-312",
+  "wide-313",
+  "wide-314",
+  "wide-315",
+  "wide-316",
+  "wide-317",
+  "wide-318",
+  "wide-319",
+  "wide-320"
+]

tools/graphrag/evaluation/results/v7_preparation/source_corrections.json

@@ -0,0 +1,151 @@
+{
+  "questions": [
+    {
+      "id": "regression-0597",
+      "old_bindings": [
+        {
+          "person": "Stanislav Matsunych",
+          "start_year": 2017
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-dev-0630",
+      "old_bindings": [
+        {
+          "person": "Jakub Maruniak",
+          "start_year": 2016
+        },
+        {
+          "person": "Márk Fehér",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-dev-0631",
+      "old_bindings": [
+        {
+          "person": "Dominik Nagy",
+          "start_year": 2016
+        },
+        {
+          "person": "Patrik Pavlišin",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-083",
+      "old_bindings": [
+        {
+          "person": "Lukáš Pokrývka",
+          "start_year": 2016
+        },
+        {
+          "person": "Maroš Harahus",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-088",
+      "old_bindings": [
+        {
+          "person": "Ján Holp",
+          "start_year": 2016
+        },
+        {
+          "person": "Maroš Harahus",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-090",
+      "old_bindings": [
+        {
+          "person": "Martin Jancura",
+          "start_year": 2017
+        },
+        {
+          "person": "Stanislav Matsunych",
+          "start_year": 2017
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-105",
+      "old_bindings": [
+        {
+          "person": "Dominik Nagy",
+          "start_year": 2016
+        },
+        {
+          "person": "Maroš Harahus",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-107",
+      "old_bindings": [
+        {
+          "person": "Dominik Nagy",
+          "start_year": 2016
+        },
+        {
+          "person": "Márk Fehér",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-111",
+      "old_bindings": [
+        {
+          "person": "Dominik Nagy",
+          "start_year": 2016
+        },
+        {
+          "person": "Tomáš Kuchárik",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    },
+    {
+      "id": "wide-112",
+      "old_bindings": [
+        {
+          "person": "Jakub Maruniak",
+          "start_year": 2016
+        },
+        {
+          "person": "Tomáš Kuchárik",
+          "start_year": 2016
+        }
+      ],
+      "new_status": "no_answer",
+      "reason": "Directory cohort is not a documented study-start year."
+    }
+  ],
+  "policy": "Only explicit study-start text is a fact. Cohort remains separate metadata; missing means unknown."
+}

tools/graphrag/evaluation/results/v7_preparation/source_review.csv

@@ -0,0 +1,21 @@
+id,category,decision,reviewer,notes,sources,human_review
+fresh7-060,candidate_set_aggregation,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2016/darius_lindvai/README.md; pages/students/2019/alina_vitko/README.md,not_performed
+fresh7-066,candidate_set_aggregation,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2018/adrian_remias/README.md; pages/students/2019/patrik_pokrivcak/README.md,not_performed
+fresh7-067,conjunctive_disambiguation,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2022/oleh_poiasnik/README.md,not_performed
+fresh7-072,conjunctive_disambiguation,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2019/alina_vitko/README.md,not_performed
+fresh7-051,document_author_category_multi_hop,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2016/darius_lindvai/README.md,not_performed
+fresh7-052,document_author_category_multi_hop,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2020/vladyslav_krupko/README.md,not_performed
+fresh7-073,negative_lookup,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2017/stanislav_matsunych/README.md,not_performed
+fresh7-080,negative_lookup,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2021/oleksandr_hryshchenko/README.md,not_performed
+fresh7-041,person_keyword_title,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2021/artur_hyrenko/README.md,not_performed
+fresh7-050,person_keyword_title,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2019/kristian_sopkovic/README.md,not_performed
+fresh7-053,person_multiple_categories,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2020/pavol_hudak/README.md,not_performed
+fresh7-059,person_multiple_categories,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2018/adrian_remias/README.md,not_performed
+fresh7-031,person_title_year,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2016/patrik_pavlisin/README.md,not_performed
+fresh7-040,person_title_year,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2022/jan_ptak/README.md,not_performed
+fresh7-011,same_academic_year_pair,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2018/adrian_remias/README.md; pages/students/2018/david_omasta/README.md,not_performed
+fresh7-020,same_academic_year_pair,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2021/oleksandr_hryshchenko/README.md; pages/students/2022/jan_ptak/README.md,not_performed
+fresh7-021,same_start_year_pair,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2020/david_kostilnik/README.md; pages/students/2020/kamil_tomcufcik/README.md,not_performed
+fresh7-030,same_start_year_pair,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2021/artur_hyrenko/README.md; pages/students/2021/matej_novotny/README.md,not_performed
+fresh7-001,shared_keyword_pair,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/interns/cesar_gutierrez/README.md; pages/students/2022/jakub_schwarc/README.md,not_performed
+fresh7-010,shared_keyword_pair,supported_by_raw_markdown,AI direct raw-source review; not human,Literal taxonomy and explicit academic/study-start evidence checked; no directory inference.,pages/students/2021/eduard_matovka/README.md; pages/students/2021/vladimir_ferko/README.md,not_performed

tools/graphrag/evaluation/results/v7_preparation/source_review_packet.json

@@ -0,0 +1,751 @@
+[
+  {
+    "id": "fresh7-060",
+    "category": "candidate_set_aggregation",
+    "question": "Z kandidátov Alina Vitko, Dárius Lindvai, Adrián Remiáš vyber osoby s prácou v roku 2021. Koľko ich je?",
+    "expected_bindings": [
+      {
+        "person": "Alina Vitko",
+        "year": 2021
+      },
+      {
+        "person": "Dárius Lindvai",
+        "year": 2021
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2019/alina_vitko/README.md",
+        "proof": {
+          "document_path": "pages/students/2019/alina_vitko/README.md",
+          "sha256": "9b071fe2aff470c275ae027be236fe836f741f774cabb9cbbb55304a8d5102cd",
+          "field": "academic_record",
+          "academic_kind": "Work",
+          "academic_type": "bachelor_project",
+          "title": "Vytovrenie prehľadu existujúcich systémov QA.",
+          "year": 2021,
+          "heading_line": 70,
+          "title_line": 72,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Alina Vitko\n3: published: true\n4: taxonomy:\n5:     category: [bp2022]\n6:     tag: [ir,qa,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n69: \n70: ## Bakalársky projekt 2021\n71: \n72: Vytovrenie prehľadu existujúcich systémov QA.\n73: "
+      },
+      {
+        "path": "pages/students/2016/darius_lindvai/README.md",
+        "proof": {
+          "document_path": "pages/students/2016/darius_lindvai/README.md",
+          "sha256": "fcaa47bb9708a6439eec5c4b7f7b4ed6c7359e423e1d940e1aed9f44957e6d62",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "diploma_thesis",
+          "title": "Obnovenie interpunkcie pomocou hlbokých neurónových sietí",
+          "year": 2021,
+          "heading_line": null,
+          "title_line": 14,
+          "year_basis": "frontmatter_category"
+        },
+        "raw_excerpt": "1: ---\n2: title: Dárius Lindvai\n3: published: true\n4: taxonomy:\n5:     category: [dp2021,bp2019]\n6:     tag: [nn,interpunction,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: # Dárius Lindvai\n13: \n14: Názov: Obnovenie interpunkcie pomocou hlbokých neurónových sietí\n15: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-066",
+    "category": "candidate_set_aggregation",
+    "question": "Z kandidátov Adrián Remiáš, Patrik Pokrivčák, Alina Vitko vyber osoby s prácou v roku 2024. Koľko ich je?",
+    "expected_bindings": [
+      {
+        "person": "Adrián Remiáš",
+        "year": 2024
+      },
+      {
+        "person": "Patrik Pokrivčák",
+        "year": 2024
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2018/adrian_remias/README.md",
+        "proof": {
+          "document_path": "pages/students/2018/adrian_remias/README.md",
+          "sha256": "e5bfe1d5d4e626beaa636a02d34ec58cf2cff9b0e57833905af35b80bf1465a2",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "diploma_thesis",
+          "title": null,
+          "year": 2024,
+          "heading_line": 15,
+          "title_line": null,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Adrián Remiáš\n3: published: true\n4: date: 01-02-2014\n5: taxonomy:\n6:     category: [dp2024]\n7:     tag: [nlp, transformers, interpuction]\n8:     author: Daniel Hladek\n9: ---\n14: \n15: # Diplomová práca 2024\n16: "
+      },
+      {
+        "path": "pages/students/2019/patrik_pokrivcak/README.md",
+        "proof": {
+          "document_path": "pages/students/2019/patrik_pokrivcak/README.md",
+          "sha256": "3e3957ad0f7aca59d6eb64c38a3f5e21d8fd953200825d0d0e5cacabfbc35a7d",
+          "field": "academic_record",
+          "academic_kind": "Work",
+          "academic_type": "diploma_project",
+          "title": null,
+          "year": 2024,
+          "heading_line": 64,
+          "title_line": null,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Patrik Pokrivčák\n3: published: true\n4: taxonomy:\n5:     category: [dp2025]\n6:     tag: [nlp, hate]\n7:     author: Daniel Hladek\n8: ---\n9: \n63: \n64: ## Diplomový projekt 2024\n65: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-067",
+    "category": "conjunctive_disambiguation",
+    "question": "Ktorá z osôb Oleh Poiasnik a Adrián Remiáš má kľúčové slovo nlp a jej práca je z roku 2025?",
+    "expected_bindings": [
+      {
+        "person": "Oleh Poiasnik",
+        "keyword": "nlp",
+        "year": 2025
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2022/oleh_poiasnik/README.md",
+        "proof": {
+          "document_path": "pages/students/2022/oleh_poiasnik/README.md",
+          "sha256": "c2f7dc3f61f8b4a5b43ded1e3ae4c7d3e7f874439ee6b17dd9c8d45dabc25bc8",
+          "field": "taxonomy.tag",
+          "value": "nlp"
+        },
+        "raw_excerpt": "1: ---\n2: title: Oleh Poiasnik\n3: published: true\n4: taxonomy:\n5:     category: [bp2025,dp2027]\n6:     tag: [rag,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2022/oleh_poiasnik/README.md",
+        "proof": {
+          "document_path": "pages/students/2022/oleh_poiasnik/README.md",
+          "sha256": "c2f7dc3f61f8b4a5b43ded1e3ae4c7d3e7f874439ee6b17dd9c8d45dabc25bc8",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "bachelor_thesis",
+          "title": "Generovanie jazyka s podporou vyhľadávania v doménovo orientovaných dátach",
+          "year": 2025,
+          "heading_line": 80,
+          "title_line": 86,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Oleh Poiasnik\n3: published: true\n4: taxonomy:\n5:     category: [bp2025,dp2027]\n6:     tag: [rag,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n79: \n80: # Bakalárska práca 2025\n81: \n85: \n86: Generovanie jazyka s podporou vyhľadávania v doménovo orientovaných dátach\n87: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-072",
+    "category": "conjunctive_disambiguation",
+    "question": "Ktorá z osôb Alina Vitko a Andrii Pervashov má kľúčové slovo nlp a jej práca je z roku 2022?",
+    "expected_bindings": [
+      {
+        "person": "Alina Vitko",
+        "keyword": "nlp",
+        "year": 2022
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2019/alina_vitko/README.md",
+        "proof": {
+          "document_path": "pages/students/2019/alina_vitko/README.md",
+          "sha256": "9b071fe2aff470c275ae027be236fe836f741f774cabb9cbbb55304a8d5102cd",
+          "field": "taxonomy.tag",
+          "value": "nlp"
+        },
+        "raw_excerpt": "1: ---\n2: title: Alina Vitko\n3: published: true\n4: taxonomy:\n5:     category: [bp2022]\n6:     tag: [ir,qa,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2019/alina_vitko/README.md",
+        "proof": {
+          "document_path": "pages/students/2019/alina_vitko/README.md",
+          "sha256": "9b071fe2aff470c275ae027be236fe836f741f774cabb9cbbb55304a8d5102cd",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "bachelor_thesis",
+          "title": "Systémy pre generovanie odpovede na otázku v prirodzenom jazyku",
+          "year": 2022,
+          "heading_line": 12,
+          "title_line": 14,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Alina Vitko\n3: published: true\n4: taxonomy:\n5:     category: [bp2022]\n6:     tag: [ir,qa,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n11: \n12: # Bakalárska práca 2022\n13: \n14: Systémy pre generovanie odpovede na otázku v prirodzenom jazyku\n15: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-051",
+    "category": "document_author_category_multi_hop",
+    "question": "Kto je autor dokumentu osoby Dárius Lindvai, ktorý patrí do kategórie dp2021?",
+    "expected_bindings": [
+      {
+        "person": "Dárius Lindvai",
+        "category": "dp2021",
+        "document_author": "Daniel Hladek",
+        "document_path": "pages/students/2016/darius_lindvai/README.md",
+        "same_document": true
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2016/darius_lindvai/README.md",
+        "proof": {
+          "document_path": "pages/students/2016/darius_lindvai/README.md",
+          "sha256": "fcaa47bb9708a6439eec5c4b7f7b4ed6c7359e423e1d940e1aed9f44957e6d62",
+          "field": "taxonomy.category",
+          "value": "dp2021"
+        },
+        "raw_excerpt": "1: ---\n2: title: Dárius Lindvai\n3: published: true\n4: taxonomy:\n5:     category: [dp2021,bp2019]\n6:     tag: [nn,interpunction,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: # Dárius Lindvai"
+      },
+      {
+        "path": "pages/students/2016/darius_lindvai/README.md",
+        "proof": {
+          "document_path": "pages/students/2016/darius_lindvai/README.md",
+          "sha256": "fcaa47bb9708a6439eec5c4b7f7b4ed6c7359e423e1d940e1aed9f44957e6d62",
+          "field": "taxonomy.author",
+          "value": "Daniel Hladek"
+        },
+        "raw_excerpt": "1: ---\n2: title: Dárius Lindvai\n3: published: true\n4: taxonomy:\n5:     category: [dp2021,bp2019]\n6:     tag: [nn,interpunction,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: # Dárius Lindvai"
+      }
+    ]
+  },
+  {
+    "id": "fresh7-052",
+    "category": "document_author_category_multi_hop",
+    "question": "Kto je autor dokumentu osoby Vladyslav Krupko, ktorý patrí do kategórie bp2024?",
+    "expected_bindings": [
+      {
+        "person": "Vladyslav Krupko",
+        "category": "bp2024",
+        "document_author": "Daniel Hladek",
+        "document_path": "pages/students/2020/vladyslav_krupko/README.md",
+        "same_document": true
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2020/vladyslav_krupko/README.md",
+        "proof": {
+          "document_path": "pages/students/2020/vladyslav_krupko/README.md",
+          "sha256": "4c28e6d38651bb2cafcbf4a18e35210cd369aef2caff6c5b0aa52fac38f0fe7d",
+          "field": "taxonomy.category",
+          "value": "bp2024"
+        },
+        "raw_excerpt": "1: ---\n2: title: Vladyslav Krupko\n3: published: true\n4: taxonomy:\n5:     category: [bp2024]\n6:     tag: [spelling]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2020/vladyslav_krupko/README.md",
+        "proof": {
+          "document_path": "pages/students/2020/vladyslav_krupko/README.md",
+          "sha256": "4c28e6d38651bb2cafcbf4a18e35210cd369aef2caff6c5b0aa52fac38f0fe7d",
+          "field": "taxonomy.author",
+          "value": "Daniel Hladek"
+        },
+        "raw_excerpt": "1: ---\n2: title: Vladyslav Krupko\n3: published: true\n4: taxonomy:\n5:     category: [bp2024]\n6:     tag: [spelling]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-073",
+    "category": "negative_lookup",
+    "question": "Aký názov práce má osoba Stanislav Matsunych v roku 2018?",
+    "expected_bindings": [],
+    "raw_sources": [
+      {
+        "path": "pages/students/2017/stanislav_matsunych/README.md",
+        "proof": {
+          "document_path": "pages/students/2017/stanislav_matsunych/README.md",
+          "sha256": "90ffa11eff4a4360b0252dbc53a116c3a5777b3d5dd86ce5c89c02bdd2051c73",
+          "field": "closed_world_metadata",
+          "tags": [
+            "nn",
+            "lm",
+            "nlp"
+          ],
+          "categories": [
+            "bp2020"
+          ]
+        },
+        "raw_excerpt": "1: ---\n2: title: Stanislav Matsunych\n3: published: true\n4: taxonomy:\n5:     category: [bp2020]\n6:     tag: [nn,lm,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: # Stanislav Matsunych"
+      }
+    ]
+  },
+  {
+    "id": "fresh7-080",
+    "category": "negative_lookup",
+    "question": "Aký názov práce má osoba Oleksandr Hryshchenko v roku 2018?",
+    "expected_bindings": [],
+    "raw_sources": [
+      {
+        "path": "pages/students/2021/oleksandr_hryshchenko/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/oleksandr_hryshchenko/README.md",
+          "sha256": "4b073992b715b6c75b79fd2c8f850a69f940966e3a61a82c625c32c945a93036",
+          "field": "closed_world_metadata",
+          "tags": [
+            "ir",
+            "nlp"
+          ],
+          "categories": [
+            "dp2026"
+          ]
+        },
+        "raw_excerpt": "1: ---\n2: title: Oleksandr Hryshchenko\n3: published: true\n4: taxonomy:\n5:     category: [dp2026]\n6:     tag: [ir,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-041",
+    "category": "person_keyword_title",
+    "question": "Aký názov práce osoby Artur Hyrenko v roku 2026 súvisí s kľúčovým slovom dialog?",
+    "expected_bindings": [
+      {
+        "person": "Artur Hyrenko",
+        "academic_kind": "Thesis",
+        "academic_type": "diploma_thesis",
+        "year": 2026,
+        "title": "Zvýšenie bezpečnosti veľkých jazykových modelov",
+        "keyword": "dialog"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2021/artur_hyrenko/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/artur_hyrenko/README.md",
+          "sha256": "0568261f4ade4de5472edcfeb1ffed9abae18dc67e4050f4ce2ad088839a2bb3",
+          "field": "taxonomy.tag",
+          "value": "dialog"
+        },
+        "raw_excerpt": "1: ---\n2: title: Artur Hyrenko\n3: published: true\n4: taxonomy:\n5:     category: [dp2026]\n6:     tag: [chatbot,rasa,dialog,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2021/artur_hyrenko/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/artur_hyrenko/README.md",
+          "sha256": "0568261f4ade4de5472edcfeb1ffed9abae18dc67e4050f4ce2ad088839a2bb3",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "diploma_thesis",
+          "title": "Zvýšenie bezpečnosti veľkých jazykových modelov",
+          "year": 2026,
+          "heading_line": 13,
+          "title_line": 15,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Artur Hyrenko\n3: published: true\n4: taxonomy:\n5:     category: [dp2026]\n6:     tag: [chatbot,rasa,dialog,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n12: \n13: # Diplomová práca 2026\n14: \n15: Zvýšenie bezpečnosti veľkých jazykových modelov\n16: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-050",
+    "category": "person_keyword_title",
+    "question": "Aký názov práce osoby Kristián Sopkovič v roku 2022 súvisí s kľúčovým slovom nlp?",
+    "expected_bindings": [
+      {
+        "person": "Kristián Sopkovič",
+        "academic_kind": "Thesis",
+        "academic_type": "bachelor_thesis",
+        "year": 2022,
+        "title": "Model Spacy pre spracovanie prirodzeného jazyka v slovenčine",
+        "keyword": "nlp"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2019/kristian_sopkovic/README.md",
+        "proof": {
+          "document_path": "pages/students/2019/kristian_sopkovic/README.md",
+          "sha256": "3286960fb30f584df35455f9b6dfae81ca4827a1d711d2bcd6e48c3f150c9279",
+          "field": "taxonomy.tag",
+          "value": "nlp"
+        },
+        "raw_excerpt": "1: ---\n2: title: Kristián Sopkovič\n3: published: true\n4: taxonomy:\n5:     category: [bp2022,dp2024]\n6:     tag: [spacy,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2019/kristian_sopkovic/README.md",
+        "proof": {
+          "document_path": "pages/students/2019/kristian_sopkovic/README.md",
+          "sha256": "3286960fb30f584df35455f9b6dfae81ca4827a1d711d2bcd6e48c3f150c9279",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "bachelor_thesis",
+          "title": "Model Spacy pre spracovanie prirodzeného jazyka v slovenčine",
+          "year": 2022,
+          "heading_line": 308,
+          "title_line": 312,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Kristián Sopkovič\n3: published: true\n4: taxonomy:\n5:     category: [bp2022,dp2024]\n6:     tag: [spacy,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n307: \n308: # Bakalárska práca 2022\n309: \n311: \n312: Model Spacy pre spracovanie prirodzeného jazyka v slovenčine\n313: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-053",
+    "category": "person_multiple_categories",
+    "question": "Vypíš všetky kategórie dokumentov osoby Pavol Hudák.",
+    "expected_bindings": [
+      {
+        "person": "Pavol Hudák",
+        "category": "dp2025"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2020/pavol_hudak/README.md",
+        "proof": {
+          "document_path": "pages/students/2020/pavol_hudak/README.md",
+          "sha256": "33d06fccc6451e5f19b27b0cad001a34e81635b4ea294a01dec7771c823abc40",
+          "field": "taxonomy.category",
+          "value": "dp2025"
+        },
+        "raw_excerpt": "1: ---\n2: title: Pavol Hudák\n3: published: true\n4: taxonomy:\n5:     category: [dp2025]\n6:     tag: [nlp,qa]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-059",
+    "category": "person_multiple_categories",
+    "question": "Vypíš všetky kategórie dokumentov osoby Adrián Remiáš.",
+    "expected_bindings": [
+      {
+        "person": "Adrián Remiáš",
+        "category": "dp2024"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2018/adrian_remias/README.md",
+        "proof": {
+          "document_path": "pages/students/2018/adrian_remias/README.md",
+          "sha256": "e5bfe1d5d4e626beaa636a02d34ec58cf2cff9b0e57833905af35b80bf1465a2",
+          "field": "taxonomy.category",
+          "value": "dp2024"
+        },
+        "raw_excerpt": "1: ---\n2: title: Adrián Remiáš\n3: published: true\n4: date: 01-02-2014\n5: taxonomy:\n6:     category: [dp2024]\n7:     tag: [nlp, transformers, interpuction]\n8:     author: Daniel Hladek\n9: ---"
+      }
+    ]
+  },
+  {
+    "id": "fresh7-031",
+    "category": "person_title_year",
+    "question": "Aký názov práce a rok práce sú pri osobe Patrik Pavlišin uvedené pre rok 2020?",
+    "expected_bindings": [
+      {
+        "person": "Patrik Pavlišin",
+        "academic_kind": "Thesis",
+        "academic_type": "bachelor_thesis",
+        "year": 2020,
+        "title": "Metódy automatického prekladu"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2016/patrik_pavlisin/README.md",
+        "proof": {
+          "document_path": "pages/students/2016/patrik_pavlisin/README.md",
+          "sha256": "dc1c648d7cd9b0f0e021c57f9b98189e37c100871439e7e304390715913f2763",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "bachelor_thesis",
+          "title": "Metódy automatického prekladu",
+          "year": 2020,
+          "heading_line": 242,
+          "title_line": 246,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Patrik Pavlišin\n3: published: true\n4: taxonomy:\n5:     category: [dp2022,bp2020,tp2021]\n6:     tag: [nmt,translation,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: # Patrik Pavlišin\n241: \n242: ## Bakalárska práca 2020\n243: \n245: \n246: Názov bakalárskej práce: Metódy automatického prekladu\n247: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-040",
+    "category": "person_title_year",
+    "question": "Aký názov diplomového projektu a rok práce sú pri osobe Ján Pták uvedené pre rok 2026?",
+    "expected_bindings": [
+      {
+        "person": "Ján Pták",
+        "academic_kind": "Work",
+        "academic_type": "diploma_project",
+        "year": 2026,
+        "title": "Agent pre manažment záverečných prác"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2022/jan_ptak/README.md",
+        "proof": {
+          "document_path": "pages/students/2022/jan_ptak/README.md",
+          "sha256": "131cf053764fd5423541ebe672c3b4f5da875cfb8e4b821f1335ddb967282282",
+          "field": "academic_record",
+          "academic_kind": "Work",
+          "academic_type": "diploma_project",
+          "title": "Agent pre manažment záverečných prác",
+          "year": 2026,
+          "heading_line": 13,
+          "title_line": 18,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Ján Pták\n3: published: true\n4: taxonomy:\n5:     category: [dp2027]\n6:     tag: [rag,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n12: \n13: # Diplomový projekt 2026\n14: \n17: \n18: Agent pre manažment záverečných prác\n19: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-011",
+    "category": "same_academic_year_pair",
+    "question": "Aké spoločné roky práce spájajú osoby Adrián Remiáš a Dávid Omasta?",
+    "expected_bindings": [
+      {
+        "person": "Adrián Remiáš",
+        "year": 2024
+      },
+      {
+        "person": "Dávid Omasta",
+        "year": 2024
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2018/adrian_remias/README.md",
+        "proof": {
+          "document_path": "pages/students/2018/adrian_remias/README.md",
+          "sha256": "e5bfe1d5d4e626beaa636a02d34ec58cf2cff9b0e57833905af35b80bf1465a2",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "diploma_thesis",
+          "title": null,
+          "year": 2024,
+          "heading_line": 15,
+          "title_line": null,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Adrián Remiáš\n3: published: true\n4: date: 01-02-2014\n5: taxonomy:\n6:     category: [dp2024]\n7:     tag: [nlp, transformers, interpuction]\n8:     author: Daniel Hladek\n9: ---\n14: \n15: # Diplomová práca 2024\n16: "
+      },
+      {
+        "path": "pages/students/2018/david_omasta/README.md",
+        "proof": {
+          "document_path": "pages/students/2018/david_omasta/README.md",
+          "sha256": "4164051870e1e1cf8949ed18a74196e9ea789dfbe4fee24bdeb736ca42848327",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "diploma_thesis",
+          "title": "Generatívne modely pre automatické odpovede na otázky v slovenskom jazyku",
+          "year": 2024,
+          "heading_line": 19,
+          "title_line": 25,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Dávid Omasta\n3: published: true\n4: taxonomy:\n5:     category: [dp2023,dp2024]\n6:     tag: [lm]\n7:     author: Daniel Hladek\n8: ---\n9: \n18: \n19: # Diplomová práca 2024\n20: \n24: \n25: Generatívne modely pre automatické odpovede na otázky v slovenskom jazyku\n26: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-020",
+    "category": "same_academic_year_pair",
+    "question": "Aké spoločné roky práce spájajú osoby Ján Pták a Oleksandr Hryshchenko?",
+    "expected_bindings": [
+      {
+        "person": "Ján Pták",
+        "year": 2026
+      },
+      {
+        "person": "Oleksandr Hryshchenko",
+        "year": 2026
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2022/jan_ptak/README.md",
+        "proof": {
+          "document_path": "pages/students/2022/jan_ptak/README.md",
+          "sha256": "131cf053764fd5423541ebe672c3b4f5da875cfb8e4b821f1335ddb967282282",
+          "field": "academic_record",
+          "academic_kind": "Work",
+          "academic_type": "diploma_project",
+          "title": "Agent pre manažment záverečných prác",
+          "year": 2026,
+          "heading_line": 13,
+          "title_line": 18,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Ján Pták\n3: published: true\n4: taxonomy:\n5:     category: [dp2027]\n6:     tag: [rag,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n12: \n13: # Diplomový projekt 2026\n14: \n17: \n18: Agent pre manažment záverečných prác\n19: "
+      },
+      {
+        "path": "pages/students/2021/oleksandr_hryshchenko/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/oleksandr_hryshchenko/README.md",
+          "sha256": "4b073992b715b6c75b79fd2c8f850a69f940966e3a61a82c625c32c945a93036",
+          "field": "academic_record",
+          "academic_kind": "Thesis",
+          "academic_type": "diploma_thesis",
+          "title": "Grafové neurónové siete pre získavanie informácií",
+          "year": 2026,
+          "heading_line": 13,
+          "title_line": 15,
+          "year_basis": "heading"
+        },
+        "raw_excerpt": "1: ---\n2: title: Oleksandr Hryshchenko\n3: published: true\n4: taxonomy:\n5:     category: [dp2026]\n6:     tag: [ir,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n12: \n13: # Diplomová práca 2026\n14:  \n15: Grafové neurónové siete pre získavanie informácií\n16: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-021",
+    "category": "same_start_year_pair",
+    "question": "Porovnaj rovnaký rok začiatku štúdia osôb Dávid Kostilník a Kamil Tomčufčík.",
+    "expected_bindings": [
+      {
+        "person": "Dávid Kostilník",
+        "start_year": 2020
+      },
+      {
+        "person": "Kamil Tomčufčík",
+        "start_year": 2020
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2020/david_kostilnik/README.md",
+        "proof": {
+          "document_path": "pages/students/2020/david_kostilnik/README.md",
+          "sha256": "2927524b75da59a64120fc364a0fc374986fa4d364d54e52c36800762c203264",
+          "field": "start_year",
+          "value": 2020,
+          "basis": "explicit_text",
+          "confidence": "documented"
+        },
+        "raw_excerpt": "1: ---\n2: title: Dávid Kostilník\n3: published: true\n4: taxonomy:\n5:     category: [dp2025]\n6:     tag: [ir]\n7:     author: Daniel Hladek\n8: ---\n9: \n10: rok začiatku štúdia: 2020"
+      },
+      {
+        "path": "pages/students/2020/kamil_tomcufcik/README.md",
+        "proof": {
+          "document_path": "pages/students/2020/kamil_tomcufcik/README.md",
+          "sha256": "9d865adf538bc6d7992ee4f4882713447da5c936fca17c43ca0f576072537048",
+          "field": "start_year",
+          "value": 2020,
+          "basis": "explicit_text",
+          "confidence": "documented"
+        },
+        "raw_excerpt": "1: ---\n2: title: Kamil Tomčufčík\n3: published: true\n4: taxonomy:\n5:     category: [dp2025]\n6:     tag: [lm]\n7:     author: Daniel Hladek\n8: ---\n9: \n10: rok začiatku štúdia: 2020"
+      }
+    ]
+  },
+  {
+    "id": "fresh7-030",
+    "category": "same_start_year_pair",
+    "question": "Porovnaj rovnaký rok začiatku štúdia osôb Artur Hyrenko a Matej Novotný.",
+    "expected_bindings": [
+      {
+        "person": "Artur Hyrenko",
+        "start_year": 2021
+      },
+      {
+        "person": "Matej Novotný",
+        "start_year": 2021
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2021/artur_hyrenko/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/artur_hyrenko/README.md",
+          "sha256": "0568261f4ade4de5472edcfeb1ffed9abae18dc67e4050f4ce2ad088839a2bb3",
+          "field": "start_year",
+          "value": 2021,
+          "basis": "explicit_text",
+          "confidence": "documented"
+        },
+        "raw_excerpt": "1: ---\n2: title: Artur Hyrenko\n3: published: true\n4: taxonomy:\n5:     category: [dp2026]\n6:     tag: [chatbot,rasa,dialog,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n11: rok začiatku štúdia: 2021"
+      },
+      {
+        "path": "pages/students/2021/matej_novotny/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/matej_novotny/README.md",
+          "sha256": "b083a7cadde532e3428dc9a57bc61456fcc5905bda91e011b96033d42bf5d9cb",
+          "field": "start_year",
+          "value": 2021,
+          "basis": "explicit_text",
+          "confidence": "documented"
+        },
+        "raw_excerpt": "1: ---\n2: title: Matej Novotný\n3: published: true\n4: taxonomy:\n5:     category: [dp2026]\n6:     tag: [nlp]\n7:     author: Daniel Hladek\n8: ---\n9: \n11: rok začiatku štúdia: 2021"
+      }
+    ]
+  },
+  {
+    "id": "fresh7-001",
+    "category": "shared_keyword_pair",
+    "question": "Over spoločné kľúčové slovo nlp pri osobách Cesar Abascal Gutierrez a Jakub Schwarc.",
+    "expected_bindings": [
+      {
+        "person": "Cesar Abascal Gutierrez",
+        "keyword": "nlp"
+      },
+      {
+        "person": "Jakub Schwarc",
+        "keyword": "nlp"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/interns/cesar_gutierrez/README.md",
+        "proof": {
+          "document_path": "pages/interns/cesar_gutierrez/README.md",
+          "sha256": "b0be60576eed28d4b1920e9acf198684519373dfc13171567247c853b3ed699e",
+          "field": "taxonomy.tag",
+          "value": "nlp"
+        },
+        "raw_excerpt": "1: ---\n2: title: Cesar Abascal Gutierrez\n3: published: true\n4: taxonomy:\n5:     category: [iaeste]\n6:     tag: [ner,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2022/jakub_schwarc/README.md",
+        "proof": {
+          "document_path": "pages/students/2022/jakub_schwarc/README.md",
+          "sha256": "f541a8852c2b0f6f9d7182e01a34c4e3beee9bfa55427a06b331638b11eb8be6",
+          "field": "taxonomy.tag",
+          "value": "nlp"
+        },
+        "raw_excerpt": "1: ---\n2: title: Jakub Schwarc\n3: published: true\n4: taxonomy:\n5:     category: [dp2027]\n6:     tag: [nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      }
+    ]
+  },
+  {
+    "id": "fresh7-010",
+    "category": "shared_keyword_pair",
+    "question": "Over spoločné kľúčové slovo dialog pri osobách Eduard Matovka a Vladimír Ferko.",
+    "expected_bindings": [
+      {
+        "person": "Eduard Matovka",
+        "keyword": "dialog"
+      },
+      {
+        "person": "Vladimír Ferko",
+        "keyword": "dialog"
+      }
+    ],
+    "raw_sources": [
+      {
+        "path": "pages/students/2021/eduard_matovka/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/eduard_matovka/README.md",
+          "sha256": "de2a7ae38ec87110e347956b37d0605a298ed2847f972961cc9b977da8221d43",
+          "field": "taxonomy.tag",
+          "value": "dialog"
+        },
+        "raw_excerpt": "1: ---\n2: title: Eduard Matovka\n3: published: true\n4: taxonomy:\n5:     category: [bp2024]\n6:     tag: [dialog,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      },
+      {
+        "path": "pages/students/2021/vladimir_ferko/README.md",
+        "proof": {
+          "document_path": "pages/students/2021/vladimir_ferko/README.md",
+          "sha256": "68e7d311896eda2171e5eec1ec3a8da90d326ed91acaace8058f68664a50d271",
+          "field": "taxonomy.tag",
+          "value": "dialog"
+        },
+        "raw_excerpt": "1: ---\n2: title: Vladimír Ferko\n3: published: true\n4: taxonomy:\n5:     category: [bp2024]\n6:     tag: [dialog,nlp]\n7:     author: Daniel Hladek\n8: ---\n9: "
+      }
+    ]
+  }
+]

tools/graphrag/evaluation/source_oracle.py

@@ -231,13 +231,15 @@ def load_source_oracle(wiki_root: Path = DEFAULT_WIKI) -> dict:
         if profile_path(path) != path or not document['title']:
             continue
         family = [doc for child, doc in documents.items() if profile_path(child) == path]
-        start = re.search(r'(?:rok\s+)?zaciat\w*\s+studi\w*\s*:?\s*(20\d{2})', norm('\n'.join(document['body_lines'])))
+        start = re.search(r'(?:rok\s+)?zaciat\w*\s+studi\w*\s*:?\s*([12]\d{3})', norm('\n'.join(document['body_lines'])))
         if not start:
             start = re.search(r'beginning of the study\s*(20\d{2})', norm('\n'.join(document['body_lines'])))
         people[document['title']] = {'person': document['title'], 'profile': path, 'documents': [d['path'] for d in family],
                                      'tags': sorted({tag for d in family for tag in d['tags']}, key=norm),
                                      'categories': sorted({cat for d in family for cat in d['categories']}, key=norm),
-                                     'start_year': int(start[1]) if start else int(path.split('/')[2]) if path.startswith('pages/students/') else None,
+                                     'start_year': int(start[1]) if start else None,
+                                     'start_year_basis': 'explicit_text' if start else 'missing',
+                                     'cohort_year': int(path.split('/')[2]) if path.startswith('pages/students/') else None,
                                      'academics': extract_academics(document),
                                      'year_conflicts': [r for r in extract_academics(document)
                                                        if r['year_basis'] == 'heading' and any(category_type(cat)[0] == r['academic_type']

tools/graphrag/graph_search.py

@@ -756,7 +756,7 @@ def extract_query_constraints(
         {
             int(value)
             for value in re.findall(
-                r"\b(20\d{2})\b",
+                r"\b(\d{4})\b",
                 normalized,
             )
         }
@@ -2450,6 +2450,15 @@ def search_graph_with_session(
     common = {"scope": GRAPH_SCOPE, "person_ids": person_ids, "limit": fetch_limit,
               "published_only": published_only}
     blocked_reasons = []
+    # A written year outside the current corpus is still a real constraint.
+    # Malformed/range notation is explicitly unsupported, never a global query.
+    explicit_year_tokens = re.findall(r'\b(?:roku|rokoch)\s+([^\s,?!]+)', clean_query, flags=re.I)
+    if any(re.match(r'\d', token) and not re.fullmatch(r'\d{4}[.]?', token) for token in explicit_year_tokens):
+        blocked_reasons.append('unsupported_year_constraint')
+    if constraints['mentions_start_year'] and any(
+        row.get('kind') == 'Person' and row.get('start_year') is None for row in resolved_entities
+    ):
+        blocked_reasons.append('missing_documented_start_year')
     if constraints["unresolved_keyword"]:
         blocked_reasons.append("unknown_keyword")
     if unknown_names:

tools/graphrag/graphrag.py

@@ -1003,9 +1003,8 @@ def extract_start_year(
     chunks: list[dict[str, Any]],
 ) -> int | None:
     pattern = re.compile(
-        r"rok\s+zaciatku\s+studia"
-        r"\s*\*?\s*:\s*\*?\s*"
-        r"(20\d{2})"
+        r"(?:(?:rok\s+)?zaciat\w*\s+studi\w*\s*\*?\s*:?\s*\*?\s*"
+        r"|beginning\s+of\s+the\s+study\s*:?\s*)([12]\d{3})"
     )
 
     for chunk in chunks:
@@ -1022,16 +1021,6 @@ def extract_start_year(
                 match.group(1)
             )
 
-    match = re.search(
-        r"pages/students/(20\d{2})/",
-        path.replace("\\", "/"),
-    )
-
-    if match:
-        return int(
-            match.group(1)
-        )
-
     return None
 
 
@@ -1685,6 +1674,9 @@ def build_graph_payload(
                 ),
             }
 
+        people[person_id]['start_year_basis'] = ('explicit_text' if people[person_id]['start_year'] is not None else 'missing')
+        people[person_id]['start_year_confidence'] = ('documented' if people[person_id]['start_year'] is not None else 'unknown')
+
         person_documents.add(
             (
                 person_id,