From 84faa121fe92df766d85fb2389964cb3ca8de4e7 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=EB=B0=95=EC=83=81=ED=98=B8=20Sangho=20Park?= Date: Sun, 7 Jun 2026 11:07:36 +0900 Subject: [PATCH] =?UTF-8?q?docs:=20resolve=20open=20questions,=20recompute?= =?UTF-8?q?=20ambiguity=20~10%=E2=86=92~5%=20(v2.3)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fold post-plan decisions into the spec and consensus plan: - Q1 deploy HW: undecided/mixed → delegate to hardware-adaptive auto-sizing - Q2 model strategy: collapse to single turbo model if P1 bench entity ≥95% - Q3 cancellation: cooperative (segment-boundary) is sufficient; no hard-kill - Q4 concurrency N: delegate to boot-time auto-sizing (AC-8 = ≤5s within auto N) Recompute clarity with the deep-interview model (Goal 0.96 / Constraint 0.95 / Success 0.95 → Total 0.954): ambiguity ~10% → ~5%. Residual is now entirely measurement/code-gated (AC-4 R-WER baseline, hybrid→single confirmation, CT2 GIL) — next lever is P1 bench, not further interview. Co-Authored-By: Claude Opus 4.8 --- .omc/plans/consensus-luke-scribe-stt-api.md | 20 ++++++++++++------ .../deep-interview-luke-scribe-stt-api.md | 21 +++++++++++++++++-- 2 files changed, 33 insertions(+), 8 deletions(-) diff --git a/.omc/plans/consensus-luke-scribe-stt-api.md b/.omc/plans/consensus-luke-scribe-stt-api.md index f7e31a0..40b427b 100644 --- a/.omc/plans/consensus-luke-scribe-stt-api.md +++ b/.omc/plans/consensus-luke-scribe-stt-api.md @@ -1,6 +1,6 @@ # Consensus Implementation Plan: luke_scribe — 로컬 STT 전사 API -- **Status:** `pending approval` (consensus **v2.2** — v2.1 합의 + CCG 외부리뷰(Codex/Gemini) 반영; §3.6 능력등급·§3.10 프로비저닝/WS/공유스토어/Colab) +- **Status:** `pending approval` (consensus **v2.3** — v2.2 + §11 Open Q 후속확정·모호도 ~5% 재산출(2026-06-07); v2.1 합의 + CCG 외부리뷰(Codex/Gemini) 반영; §3.6 능력등급·§3.10 프로비저닝/WS/공유스토어/Colab) - **Mode:** `--consensus --direct --deliberate` - **Source spec:** `.omc/specs/deep-interview-luke-scribe-stt-api.md` (ambiguity ~10%, PASSED) - **Project:** greenfield `/root/luke_scribe` @@ -220,10 +220,11 @@ luke_scribe/ --- -## 11. Open Questions (사용자 확인 권장) -1. 실제 배포가 **다중 GPU 워커**를 필요로 하나, 아니면 단일 T4/Colab 위주? (후자면 워커수 공식 위험 축소.) -2. turbo의 KO entity 보존율이 P1 bench에서 ≥95%면 **단일 모델로 단순화**할 의향이 있는지(VRAM/복잡도 절감). -3. 취소는 **협조적(세그먼트 경계)** 으로 충분한지, 즉시 hard-kill이 필요한지. +## 11. Open Questions → 후속 결정 반영 (RESOLVED · 2026-06-07) +1. **배포 타깃 HW:** **미정/혼합으로 확정** → 단일 타깃을 고정하지 않고 §3.6 하드웨어 적응형 자동산정에 위임. 워커수 공식·공유스토어는 부팅 실측으로 런타임 결정(설계 원칙 #1과 합치). *(잔여 검증: 특정 배포는 P1 `detect`/`bench`로.)* +2. **모델 전략 규칙(확정):** P1 bench에서 turbo의 entity 보존율 **≥95%(+R-WER ≤ 기준선)이면 배치도 turbo 단일모델로 통일**(기본 지향=단순화, VRAM·복잡도 절감, T2 스왑 회피). 미달 시에만 하이브리드(batch=large-v3) 유지. → 아키텍처 분기를 *측정 기반 결정규칙*으로 확정(분기 모호 제거). +3. **취소 의미론(확정):** **협조적 취소(세그먼트 경계)로 충분**. hard-kill 경로 비채택(필요 시 follow-up). §3.7a 그대로. +4. **동시성 상한 N(확정):** **자동산정 위임** → AC-8의 N은 부팅 시 하드웨어 기반 자동 산정(고정 목표 없음). AC-8 = "≤5s within auto-sized N". --- @@ -263,4 +264,11 @@ luke_scribe/ - 출처: `.omc/artifacts/ask/{codex,gemini}-20260603-095739.md`. 미채택(추후): webhook·Idempotency-Key·페이지네이션·`410`. --- -*Consensus v2.2 — `pending approval`. 실행(team/ralph/autopilot)은 사용자의 별도 명시 승인이 있어야만 진행됩니다. 승인 전 소스 수정·커밋·실행 스킬 호출 없음.* + +## 15. v2.3 Changelog — 후속 결정 반영 + 모호도 재산출 (2026-06-07) +- **Open Q1~Q4 확정(§11):** 배포HW=적응형 위임(미정/혼합), 모델=단일통일 규칙(P1 bench 게이트), 취소=협조적, 동시성 N=자동산정 위임. +- **모호도 재산출(딥인터뷰 모델 종합):** ~10%(인터뷰) → ~6%(계획 v2.2) → **~5%(후속 결정)**. Goal 0.96×0.40 + Constraint 0.95×0.30 + Success 0.95×0.30 = Total Clarity ≈**0.954** → Ambiguity ≈**0.046**. +- **잔여 모호(전부 측정·코드 시점에만 닫히는 known-unknown):** ① AC-4 R-WER 기준선(P1 bench), ② 하이브리드→단일 최종 확정(P1 bench 결과), ③ CT2 GIL→실시간 프로세스 분리(P3). 대화 레버 소진 → **다음 레버는 P1 bench 실행.** + +--- +*Consensus v2.3 — `pending approval`. 실행(team/ralph/autopilot)은 사용자의 별도 명시 승인이 있어야만 진행됩니다. 승인 전 소스 수정·커밋·실행 스킬 호출 없음.* diff --git a/.omc/specs/deep-interview-luke-scribe-stt-api.md b/.omc/specs/deep-interview-luke-scribe-stt-api.md index 476bfb9..ac2559f 100644 --- a/.omc/specs/deep-interview-luke-scribe-stt-api.md +++ b/.omc/specs/deep-interview-luke-scribe-stt-api.md @@ -6,12 +6,12 @@ ## Metadata - Interview ID: `di-luke-scribe-stt-20260602` - Rounds: 3 (스코어링) + 추가 아이디어 1 + 열린 결정 확정 1 -- Final Ambiguity Score: **~10%** (threshold 20%; 열린 결정 6건 확정 후) +- Final Ambiguity Score: **~10%** (threshold 20%; 열린 결정 6건 확정 후) → **재산출 ~5%** (계획 v2.2 종합 + Open Q 후속확정, 2026-06-07; 상세 §Clarity Re-computation) - Type: **greenfield** (빈 저장소 `luke_scribe`) - Generated: 2026-06-02 - Threshold: 0.2 / Threshold Source: `default` - Initial Context Summarized: no -- Status: **PASSED · 결정 확정 완료 · CCG 외부리뷰 반영(v2.2)** +- Status: **PASSED · 결정 확정 완료 · CCG 외부리뷰 반영(v2.2) · Open Q 후속확정·모호도 ~5% 재산출(2026-06-07)** ## Clarity Breakdown | Dimension | Score | Weight | Weighted | @@ -22,6 +22,20 @@ | **Total Clarity** | | | **0.904** | | **Ambiguity** | | | **0.096 (~10%)** | +## Clarity Re-computation (계획 v2.2 + 후속 결정 종합 · 2026-06-07) + +> 동일 모델로 재산출. 계획 v2.2가 P0×3/P1×4/P2×3 + 모호 AC 4건을 해소(특히 Success Criteria 수치화: AC-4 entity ≥95%, AC-7 RSS ±15%, OOM ≤2, 429/413, Phase Exit 게이트)했고, 후속 결정(취소=협조적, 모델 단일통일 규칙, HW·동시성=적응형 위임)으로 잔여 분기를 닫음. + +| Dimension | 기존 | 재산출 | Weight | Weighted | +|-----------|------|--------|--------|----------| +| Goal Clarity | 0.94 | **0.96** | 0.40 | 0.384 | +| Constraint Clarity | 0.90 | **0.95** | 0.30 | 0.285 | +| Success Criteria | 0.86 | **0.95** | 0.30 | 0.285 | +| **Total Clarity** | 0.904 | | | **0.954** | +| **Ambiguity** | 0.096 | | | **0.046 (~5%)** | + +**잔여 모호(전부 측정/코드 시점에만 닫힘):** ① AC-4 R-WER 기준선(P1 bench), ② 하이브리드→단일 최종 확정(P1 bench 결과), ③ CT2 GIL→실시간 프로세스 분리(P3). → 대화 레버 소진, 다음 레버는 **P1 bench 실행**. + --- ## Topology (확정 컴포넌트) @@ -330,6 +344,9 @@ Python 3.11+, **FastAPI** + uvicorn, **faster-whisper(CTranslate2)** (turbo + la | 4 | 결과 보관 | **7일**(설정화·자동삭제). 원본 오디오는 전사 직후 삭제. | | 5 | 파일 상한 | 모든 입력 **비동기 Job 기본**, 절대 상한 **4시간 / 2GB**(초과 `413`, 설정화). | | 6 | 화자 분리 | **옵션 포함**(pyannote, HF 토큰), 기본 off, 요청 시 `diarize=true`. | +| 7 | 취소 의미론 | **협조적 취소(세그먼트 경계)로 충분** 확정. hard-kill 비채택(필요 시 follow-up). *(2026-06-07)* | +| 8 | 모델 단순화 규칙 | P1 bench에서 turbo entity **≥95%(+R-WER≤기준선) 시 배치도 turbo 단일모델로 통일**(기본 지향=단순화). 미달 시 하이브리드 유지. *(2026-06-07)* | +| 9 | 배포HW·동시성 | **단일 타깃 미고정** → §6 하드웨어 적응형 자동산정에 위임. WS 동시 N도 자동산정(고정 목표 없음). *(2026-06-07)* | ---