**Достигнутый результат на $0 budget без fine-tuning:** **94.0% на published BIRD (v31), выше human-expert baseline 92.96% (+1.04pp).** Историческое 92.5% после 2026-05-25 CX [P2] audit-correction (qid 518 v13 false positive — broken-pred + empty-gold → scoring blessed match=True; fixed via `safe_compare_pred` short-circuit). BIRD-official set scoring; v29 = v28 + P3.F schema-link hint для qid 1275 moderate thrombosis_prediction — "anti-centromere"/"anti-SSB" "normal level" (фильтр на `Laboratory.CENTROMEA IN ('negative','0')` и `Laboratory.SSB IN ('negative','0')` через `Patient ⋈ Laboratory` на `.ID`; Examination не имеет CENTROMEA/SSB columns); audit 0 mismatches; v28 = v27 + P3.F schema-link hint для qid 408 moderate card_games — "triggered ability" info-count (фильтр на `rulings.text` через `cards.uuid = rulings.uuid` + `COUNT(DISTINCT cards.id)`); audit 0 mismatches; v27 = v26 + два P3.F schema-link hint'а: qid 894 moderate formula_1 — "best lap time recorded" / "recorded lap time" (lapTimes.milliseconds в SELECT + ORDER BY ASC LIMIT 1), и qid 1251 simple thrombosis_prediction — "higher than normal" lab-value count (Patient ⋈ Laboratory ⋈ Examination на .ID); audit 0 mismatches; v26 = v25 + P3.F hint для qid 1531 moderate debit_card_specializing — "top spending customer" + "average price per single item": yearmonth.Consumption subquery + SUM(Price/Amount) row-wise; v25 = v24 + P3.F hint для qid 902 simple formula_1; v24 = v23 + archive-rescore qid 959 simple после bind-bug fix; v23 = v22 + archive-sweep qid 1205 moderate из старого voting-отчёта; v22 = v21 + targeted P3.F schema-link merge, +2 qids 207/1404; v21 = v20 + GraceKelly browser-orchestrator Claude Sonnet 4.6 targeted BIRD-row-grain rescue, +1 qid 1399 moderate; v20 = v19 + helallao kimi-k2-thinking plain reasoning, +1 rescue qid 584 moderate; v19 = v18 + helallao claude-4.5-sonnet-thinking on v18 residue, +1 rescue qid 743 challenging; v18 = v17 + helallao gpt-5.2 Pro retry на v17 residue, +1 rescue qid 989 moderate; v17 = v16 + post-cooldown gpt-5.2-thinking+DAC, +1 rescue qid 896 challenging); 74.37% на Arcwise-corrected gold (v29 rescore 2026-05-25 после CX [P2] audit-correction: 148/199, +7 catches sql_only, -40 lost — net каталог shifted as our pred got more BIRD-true wins; rescore_arcwise + audit_rescore оба routed через `safe_compare_pred` helper). **Выше #1 paid system** AskData+GPT-4o (81.95%) на +12.05pp, и всех published free-tier no-FT (Arctic-32B 71.83%, CSC-SQL 73.67%, XiYan 75.63%). **Выше human expert baseline 92.96%** (BIRD paper) на +1.04pp. Основные рычаги: grounded-critique directed retry, Sonnet 4.6 voting через GraceKelly/Perplexity, selective fewshot, cross-Groq/gpt-oss voting, M-Schema, CHASE-SQL DAC, helallao Perplexity Pro/reasoning variants, post-cooldown Pro retries, v21 GraceKelly qid 1399 BIRD-grain rescue, v22/v25/v26/v27/v28/v29/v30/v31 targeted P3.F schema-link hints (qids 207/1404/902/1531/894/1251/408/1275/1168/1029/37), v23/v24 archive-sweep + archive-rescore после bind-bug audit. Full historical trace lives in `docs/SESSION_HANDOFF.md`.
0 commit comments