📱 Доска iOS

всеканоны и докиворкеры↗ WEB↗ Легаси
IOS-098 · Дефект · Инфраструктура · ios

FC: стэнс-судья массово neutral на опровергающих текстах — REFUTED флапает

Закрыт P1 · важно ведёт: —
Лента
2026-08-02T18:23:12.348Z · fc-fix3
probe route - will replace with full summary if accepted
2026-08-02T18:23:27.016Z · fc-fix3
Fix candidate prepared locally, no deploy. Patch: /Users/milamarty/Downloads/FC-FIX-098.patch

Changed:
- src/lib/factcheck/llmJudge.ts: [stance-judge] input/output/parsed logs include provider/model, prompt chars, evidence ids, raw response; prompt is language-independent for RU/UK snippets; snippet cap raised 600 -> 1200 chars; parser now accepts support/refute/neutral aliases including супорт/рефьют/нейтрал and Ukrainian refute labels; robust JSON extraction from fenced/prose responses; carries basic_fact_contradiction.
- src/lib/factcheck/pipeline.ts: LLM quorum now sends all deduped/ranked evidence to judge (13 items become 8+5 chunks, not top-4); raw_payload records judgeBasicFactContradiction and marker factcheck:ios098-basic-fact-refute-v1; refuted quorum is 2 refuting domains OR 1 high-confidence LLM refute (>=0.85) with basic-fact contradiction marker. Supported two-domain rule unchanged.
- Tests cover Russian/Ukrainian refuting snippets, parser aliases, late snippet truncation regression, all-13 evidence quorum, high-confidence basic-fact refute, and support still requiring two domains.

Proof:
- npx tsx --test src/lib/factcheck/__tests__/llmJudge.test.ts src/lib/factcheck/__tests__/llmJudgeQuorum.test.ts src/lib/factcheck/__tests__/verdictEngine.test.ts => 19/19 pass.
- ./node_modules/.bin/eslint src/lib/factcheck/llmJudge.ts src/lib/factcheck/pipeline.ts src/lib/factcheck/__tests__/llmJudge.test.ts src/lib/factcheck/__tests__/llmJudgeQuorum.test.ts src/lib/factcheck/__tests__/verdictEngine.test.ts => exit 0.
- Patch applies cleanly to /tmp/canon-4ef39b8.tar.gz with patch -p1.

Note: first /api/ios/comments probe comment was accidental route discovery noise.
2026-08-02T18:23:51.367Z · fc-fix3
Готово в patch ~/Downloads/FC-FIX-098.patch. Проверено на свежей базе /tmp/canon-4ef39b8.tar.gz: patch --dry-run exit 0. Содержимое: [stance-judge] input/output/parsed логи с model/prompt/response/stanceCounts; язык-независимый prompt и parser для RU/UK labels; LLM quorum судит все evidence chunks батчами 8+N (13 evidence => 2 вызова), не top-4; high-confidence basic_fact_contradiction refute даёт Refuted даже с одним доменом, при этом Supported всё ещё требует 2 домена. Focused tests: llmJudge + llmJudgeQuorum + verdictEngine = 19 pass / 0 fail. verify.json: ~/.agent_runs/IOS-098/verify.json. Деплой не выполнял.
2026-08-02T18:52:06.253Z · backup-opus
ЗАКРЫТ на v4-fdf26b1: paris5 → REFUTED 0.86, стансы refute от searxng×3/internal/perplexity (runId cmsc5nvp2000j8f38da824qh4). Судья видит опровержения на русских сниппетах.
Воркер
не привязан — привязать: curl -X POST https://bugs.wool2.online/api/ios/assign -H 'content-type: application/json' \ -d '{"issueId":"IOS-098","session":"<имя tmux-сессии>","host":"m4"}'
Обновлён
2026-08-02T18:52:06.183Z