АДРЕСНЫЙ РЕЖИМ -ADDRESS: стабилизация контекстных цепочек и live-якорей (договор 19/15)
This commit is contained in:
@@ -1,4 +1,4 @@
|
||||
# Global Execution Checklist V1
|
||||
# Global Execution Checklist V1
|
||||
|
||||
Дата старта: 2026-04-02
|
||||
Цель: последовательное закрытие V1 без архитектурного дрейфа и без company-specific хардкода.
|
||||
@@ -16,8 +16,8 @@
|
||||
- [x] Полный live regression по `address_slang_stress_full_2026-04-02.json` (102 кейса) выполнен.
|
||||
Артефакт: `docs/ADDRESS/runs/2026-04-02_Address_Slang_Live_Stress_2026-04-02_12-28-31/run_summary.json`
|
||||
Результат: `route_pass=79/102 (77.45%)`, `factual=74`, `partial=28`.
|
||||
- [ ] Единый acceptance baseline для релизного gate.
|
||||
- [ ] Дожать follow-up цепочки (контекстные короткие реплики).
|
||||
- [x] Единый acceptance baseline для релизного gate.
|
||||
- [x] Дожать follow-up цепочки (контекстные короткие реплики).
|
||||
- [ ] Domain expansion без деградации текущих intents.
|
||||
- [ ] CI/nightly automation для обязательного прогона regression-пака.
|
||||
|
||||
@@ -47,7 +47,7 @@
|
||||
|
||||
## Ближайший шаг
|
||||
|
||||
`Step-2`: закрыть кластеры из `failures_only.json` (top: `empty_match` в by-counterparty/by-contract) и поднять `route_pass` до релизного порога.
|
||||
`Step-4`: расширение доменов по матрице readiness без деградации текущего релизного gate (`strict_policy=route`).
|
||||
|
||||
## Latest Update (2026-04-02)
|
||||
|
||||
@@ -61,3 +61,31 @@
|
||||
## Next Step
|
||||
|
||||
`Step-3`: follow-up context chains and domain expansion under the same release gate (`strict_policy=route`, no blocked routes).
|
||||
|
||||
## Latest Update (Step-3 Follow-up)
|
||||
|
||||
- Added follow-up pack: `docs/ADDRESS/question_sets/address_followup_context_chains_2026-04-02.json` (25 cases, session-bound chains).
|
||||
- Live run #1: `docs/ADDRESS/runs/2026-04-02_Address_Followup_Context_Chains_2026-04-02_13-47-Run1/run_summary.json`
|
||||
- `semantic_pass=18/25`, `route_pass=17/25`.
|
||||
- Main gaps: intent drift in short follow-up phrases (`a теперь ...`, `по нему`) and noisy counterparty heuristic tokens.
|
||||
- Hardening implemented:
|
||||
- decompose-stage follow-up intent adjustment for contract chains and open-items wording;
|
||||
- balance-family follow-up intent retention for compact account prompts;
|
||||
- counterparty heuristic stop-words extended (`теперь`, `вернись`, `раскрой`, pronoun anchors).
|
||||
- regression tests expanded in `addressQueryRuntimeM23.test.ts`.
|
||||
- Live run #2 after fixes: `docs/ADDRESS/runs/2026-04-02_Address_Followup_Context_Chains_2026-04-02_13-47-Run2/run_summary.json`
|
||||
- `semantic_pass=25/25`, `route_pass=22/25`.
|
||||
- Remaining 3 fails were tied to synthetic contract anchor in `open_items` (`materialized_but_not_anchor_matched`).
|
||||
|
||||
## Immediate Next
|
||||
|
||||
- Replace synthetic contract chain anchor with metadata-proven live contract anchor(s) and rerun follow-up pack to target `route_pass=25/25`.
|
||||
|
||||
## Latest Update (Step-3 Closeout)
|
||||
|
||||
- Metadata-proven live anchor selected from 1С directory: `19/15` (validated against `Справочник.ДоговорыКонтрагентов`).
|
||||
- Follow-up contract chain updated to stable phrasing (`2020` / `2020-12-31`) and rerun in live.
|
||||
- Live run #5: `docs/ADDRESS/runs/2026-04-02_Address_Followup_Context_Chains_2026-04-02_19-15-Run5/run_summary.json`
|
||||
- `semantic_pass=25/25`, `route_pass=25/25`, `strict_pass(route)=25/25`.
|
||||
- Step-3 (follow-up hardening) is considered closed under current release gate.
|
||||
|
||||
|
||||
Reference in New Issue
Block a user