| name | nuri-siege-audit |
| description | SIEGE predictivity measurement (E4-0b v2) audit 재실행 + 해석 절차. Use when re-running 60-month predictivity audit, interpreting gate-level Δ results, or adding new variant templates. audit route 는 2026-04-22 currently closed state — 재활성화 조건 확인용. |
| disable-model-invocation | true |
| allowed-tools | Bash(ssh *) Bash(.venv/bin/python *) |
SIEGE Predictivity Audit — Closed Route 재실행 프로토콜
Source: docs/STRATEGY.md §3.8 canonical (60-month trajectory authoritative). 본 skill 은 재실행 / 해석 / 재활성화 조건.
재현 명령
ssh $DEV2_HOST 'cd ~/workspace/nuri-quant && .venv/bin/python scripts/analysis/siege_predictivity_audit.py --months 60 --save'
.venv/bin/python scripts/analysis/siege_predictivity_audit.py --months 36 --save
Artifact: data/reports/YYYY-MM-DD/e4_0b_siege_predictivity.md (gitignored). 다른 머신/run 과 diverge 시 STRATEGY §3.8 본문 숫자가 reference.
v2 Methodology
Variant ladder (4 templates × N months): momentum_top10 / equal_weight_sample / sector_concentrated / concentrated_top5 — 각 gate fire/not-fire 양쪽 sample 확보.
Gate eligibility matrix (codex Biggest Risk fix):
| 분류 | Gates | 상태 |
|---|
auditable_now | position_limit, sector_limit, leverage_ban | snapshot-native, 측정 대상 |
audit_incoherent | data_fresh_, external_data_, volatility_gate_*, drift_safe, macro_event_alignment, conflict_free | DB state 의존, snapshot coherence 없음, skip |
requires_replayed_state | stop_loss, rules_loaded | historical pnl/metadata 부재, 측정 불가 |
Hybrid metrics: Binary Δ (fired − not_fired mean fwd return + 95% bootstrap CI) primary + continuous severity OLS slope secondary.
Acceptance (codex Q5 — CI upper bound):
primary_keep: 30d CI_high < 0 AND 60d point < 0
strong_keep: 30d + 60d 모두 CI_high < 0
60-month Gate-level 결과 (2026-04-22 production)
| Gate | Fire/Not | Δ30d CI | Δ60d CI | Primary_keep |
|---|
position_limit | 47/94 | [-3.43, +8.77] | [-6.79, +12.85] | ❌ |
sector_limit | 141/0 | non-fire 부재 | — | N/A |
leverage_ban | 0/141 | fire 부재 | — | N/A |
핵심 finding: position_limit point estimate 양수 (fired > not-fired) 안정적 — §3.7 downside-predictive 가설과 반대 sign. CI 0 가로지름이나 wrong-sign 은 sample size 로 해결 안 됨 — point 가 음수 flip 해야 CI_upper<0 가능.
Prudential vs predictive 축 분리 (필수 인용 규칙)
position_limit / sector_limit / leverage_ban 는 prudential portfolio constraints + user-preference defaults 로 유효 — 사용자 감정 통제, STRATEGY §7 자동 매매 deferred, O'Neil/Minervini lineage.
그러나 synthetic audit 은 forward-downside-predictive gate 로 기능한다는 주장 미증명. rule 은 prudential 근거로만 인용, downside-predictive framing 은 §3.8 에서 tentatively refuted 로 인용.
재활성화 조건
이 route 재실행은 의미 없음. 다음 중 하나 충족 시 새 Plan consult + STRATEGY 개정:
- Actual-portfolio replay — tracker outcomes 누적 후 real-history 기반 측정
- §2.6 Symmetric amplifier 이후 upside-measurement 전용 경로 설계
- audit design 근본 redesign — variant ladder × momentum snapshot 외 source
v1 → v2 narrative + 재설계 이유 (#417 closure): gh pr view 443 / codex-reviews/PRe4-0b-v2-roundplan-20260421T175538Z.md.