SUPREME EDGE ENHANCEMENT

Source inputs (what fed this master plan)

Per-class action items (specific, not generic)

Cross-cutting P0 cluster (next 24h, all-class impact)

P1 cluster (week 1, structural)

1. Implement Codex's readiness.by_class payload contract (class state-machine fields + capped_vs_raw_pnl_gap + single_symbol_concentration + leaders.by_class + draggers.by_class).
2. Add walk-forward coverage for COMMODITY + BOND in alpha_engine/walkforward_validator.py; surface in audit_trail/dashboard_generator.py.
3. Fix drift detector — hf_stats.concept_drift.KS_D uncomputed-zero bug + refresh 19-day stale hf_stats. Wire drift → auto-pause sizing when D > 0.10 (peer remaining task "Drift-pause activation Phase 1").
4. Reconcile /audit threshold text with docs/PERFORMANCE_CHARTER.md v1.0.
5. Add last_signal_date to systems payload (Claude Code — absent for all top-6 winners).
6. Peer high-priority backlog: v3b LLM-driven signal translator (peer Phase H §A) — likely flips several NO_EDGE → MIXED/GO once spec-faithful signals replace SMA proxy. ~$1/run dispatcher.
7. Peer high-priority backlog: Re-fire P5 swarms with v3a numbers (~$0.35) — current P5 verdicts cached from pre-v3a stub numbers.
8. Investigate-then-quarantine remaining top draggers (mercury2_fast, ml_bg_system_b, copy_trader_highscore) per docs/MUTATION_THREE_AXIS_PROTOCOL.md.

P2 cluster (weeks 2-4, per-class rehab in parallel)

See per-class action lists above. Cross-cutting:

1. E-D3 — dry_run kwarg on smart_picks_engine + production_scanner + dashboard_generator (peer remaining backlog medium).
2. CPCV upgrade — swap walk-forward for CPCV in p3_backtest_runner.py (peer remaining backlog; closes project_cpcv_gap_2026_04_28).
3. Per-asset-class deep-dive swarm questions — 35 specific questions across 7 classes captured in DAILY_IDEAS.MD §B (peer remaining backlog).
4. Re-run cross-permutations after data-trust + walk-forward fixes; verdict on per-class edge existence (Claude Code).
5. Audit HC filter against confidence-inversion on ETF/CRYPTO per memory project_performance_reality.

P3 / P4 / P5 (longer horizon)

• P3: Decay-replacement pipeline (peer remaining backlog low); HEAD-check rate-limiter; cross-link research_index.html ↔ edge_stability.html; NO_EDGE knowledge base; tests for tools/research/ (peer remaining backlog low). Riskfolio-Lib status: 9 files killed by peer Phase D (dep no-install on Py 3.14 + duplicates pypfopt fallback). Reconsider only if Py-version path opens.
• P4: FRED macro-filter wire-up (FRED_API_KEY to secrets, gate strategies on regime); Kalshi pairwise consensus with Polymarket (pm_consensus_overlay.py sidecar).
• P5: Pilot paper-trade real-time variance test only AFTER class state machine reaches SHADOW for all six major classes (Codex all-classes-first).

Open PR triage (24 open as of 2026-05-11 21:00 UTC)

Recommended PR merge order (master-plan P0 priority):

1. #904 (research orchestrator + edge stability) — unblocks verdicts
2. #876 (FOREX pnl_pct clamp) — fixes FOREX unit corruption before FOREX cap goes in
3. #891 + #884 (mysql_sync category + closed_at fixes) — truth-layer for Codex's data-trust gate
4. #887 (WIN_RATE_TRAP_BLACKLIST) — overlaps with P0 #2 + #3; verify no duplicate blocks
5. #902 (b13 regime filter) — enables Codex REHAB→OOS_READY transition; investigate the 1 failed check first
6. #892 (safe_db_archive Hermes gate) — gates any subsequent DB write
7. #893 (orphan_resolver_dryrun) — read-only preview, aligns with E-D3
8. #846 (Shadow Probation panel) — UI for state-machine visualization
9. #878 (short_engine BULL regime gate) — CRYPTO regime filter
10. #885 (risk_policy v2 crypto caps) — CRYPTO de-concentration

Real-money gate (master) — synthesis

Adopt Codex's class state machine as governance scaffold, with Kimi's 7-check Go/No-Go as per-class checklist, Cursor's Tier-2 measurable criteria as the numerical floor, and Copilot's 2-consecutive-weekly as the persistence test.

Class state machine

• BLOCKED — any global gate red (stale payload, DB-health red, drift_alert, walk-forward missing)
• REHAB — global gates green; class still under quarantine/feature-add
• OOS_READY — n ≥ 100 AND Tier-2 AND positive OOS Sharpe AND OOS consistency ≥ 80% AND concentration caps OK
• SHADOW — 14-30d forward shadow with realized-vs-expected variance inside tolerance band
• LIVE_ELIGIBLE — all six major classes (CRYPTO/EQUITY/ETF/FOREX/COMMODITY/BOND) at SHADOW simultaneously; truth-layer green for 7 consecutive days; user personally approves

Per-class current state (master-plan declaration)

Current LIVE_ELIGIBLE count: 0/6. Earliest LIVE_ELIGIBLE target: not before week 8 given REHAB→OOS_READY (2-4 weeks) + SHADOW (2-4 weeks) for the fastest class.

Test plan (per-class + cross-cutting)

Cross-cutting

• Re-run money-maker-ready skill on freshest deployed payload after every pipeline change; require class-labeled claims in every report.
• Contract tests for readiness.by_class, leaders.by_class, draggers.by_class, concentration/caveat fields (Codex).
• Regression test: fail CI if dashboard threshold text diverges from docs/PERFORMANCE_CHARTER.md.
• Integration test: red DB-health or active drift forces BLOCKED in state machine.
• Shadow-tracking acceptance test: variance ≤ agreed tolerance band before LIVE_ELIGIBLE.
• Existing tests verified by peer Phase C: 94/94 unit tests for blocklist enforcement (tests/test_quarantine_unit_blocklist.py); 16 smoke tests for edge stability (tests/test_edge_stability_smoke.py) via peer Phase I commit a9e045a757f.

Per-class

• COMMODITY: walk-forward.by_class[COMMODITY].folds > 0; concentration disclosure renders; multi_asset_cot DB-verified PF within 2pp of systems payload.
• EQUITY: bottom-symbol pruning improves recent PF without OOS consistency drop < 80%; capped MDD verified; claude_gainer_st blocklist enforcement test (intake + exec gates).
• CRYPTO: next fwd_vs_bt_divergence build shows ≤ 2 baby_strats flags (down from 12); kimi_signal_tracking generates 0 new picks in test env; quan_engine volume share ≤ 12% in next 7-day window.
• FOREX: PR #876 clamp test passes; OOS Sharpe non-negative on post-fix 30d window before any sizing un-cap.
• ETF: universe expansion adds ≥ 3 leveraged-ETF-free new symbols; n reaches 120 within 30d.
• BOND: walk-forward.by_class[BOND] emits ≥ 5 folds; promotion only at n ≥ 100.
• FUTURES: document why module is silent; either ship re-emission plan or formally retire from /audit.

Peer plan v2 (swarm-revised) — already shipped

Peer Claude-B session pushed plan v2 + production code mid-document. Tracked here so this master plan stays current.

Plan v2 swarm-revised changes (commit 57d267a28e6)

1. Phase 4 ↔ Phase 2 reorder (unanimous swarm) — measurement infrastructure before per-class scaling.
2. NEW Phase 1.5 drift-clearance gate — must hold drift_alert=false for 7 consecutive days before Phase 2 advances.
3. Walk-forward by_class downgraded to ADVISORY while drift hot — its numbers are themselves drifting in real-time during regime collapse. Phase 1.5.3 wire-in tags cards advisory_only=true; demotions deferred.
4. 3 baseline numbers corrected by red-team:
   • EQUITY consistency: 87.5% (not 75% — Cursor's original cite was 12.5pp understated)
   • CRYPTO consistency: 68% (not 84% — original was 16pp overstated)
   • FOREX consistency: 48.1% (1.8pp correction, ≈ unchanged)
   Red-team read at 2026-05-11T21Z snapshot vs claude-b's read at 2026-05-10T04Z snapshot — discrepancy implies walk-forward.by_class IS drifting; argues for snapshotting at known cutoffs not live-reading.
5. 3 risks added to plan v2 risk register:
   • Regime-overfit — Tier-1 promotion gate trained on a single high-VIX regime period may collapse when regime shifts.
   • No rollback — Opt B Tier-1 demotion has no rollback ledger; if a wrong demotion happens, no easy undo.
   • Plan cites non-existent funcs — walk_forward_by_strategy() is PROPOSED-NEW, does NOT exist yet. Master plan must not assume it.

Production code shipped this session (peer)

Concept-drift root cause (peer T4 verdict)

VIX -44.64% / 30d real regime collapse since 2026-04-22 — confirmed not pipeline noise. KS_D 0.31 vs 0.047 critical. Source: reports/concept_drift_root_cause_2026-05-11.md. This validates Codex's "fix truth layer first" position: drift is real and regime-driven, not a metrics artifact.

Master plan ripple-effects

• Re-rank P1 #3 (drift detector fix): KS_D is computed; the drift is REAL. Action shifts from "fix KS_D uncomputed-zero" to "wire 7-consecutive-day drift_alert=false check before Phase 2 advances" (Phase 1.5.1).
• De-rank walk-forward.by_class trust while drift hot. Per-class promotion gates above should use edge_stability sidecar as primary verdict, walk-forward.by_class as advisory.
• Add Phase 1.5.1 + 4.1 + 4.2 to P1:
  • 1.5.1: 7-consecutive-day drift_alert=false history check (needs daily snapshot scaffolding)
  • 4.1: DRIFT_AUTO_PAUSE_DRY_RUN=1 env-gate for 7d "would-pause" logging
  • 4.2: class_capital_gate(asset_class) → (allow_size, max_pct, reason) with capital_gate_log.jsonl trail
• EQUITY edge stronger than thought: consistency 87.5% (not 75%). Strengthens EQUITY's case for first promotion.
• CRYPTO edge weaker than thought: consistency 68% (not 84%). CRYPTO promotion gate now requires consistency lift of ≥ 12pp post-quarantine, not 0pp.

Buffy enhancements (deepseek-v4-pro, PENDING) — queue into P1

Buffy session at docs/chatlogs/chatlog_2026-05-11_buffy_review.md + docs/chatlogs/progress_2026-05-11_buffy_enhancements.md. Code review of Opt-A / Opt-B / W4 found 3 issues + 4 enhancement opportunities. All status PENDING (not yet committed).

Buffy methodology note: each enhancement follows before/change/after/verify/evidence pattern with line-range citations + swarm cross-check round + typecheck + code review log. Three rows in tracking tables currently empty — work hasn't started yet.

Issues found but not yet enhanced (queue):

• benchmark_return('FOREX') returns None → E1 fix
• Per-pick disk I/O on hot path → E2 fix
• No staging mode for drift pause → E5 fix

Action: Buffy to proceed with E1/E2/E4/E5. Each lands as separate small PR with co-author + commit message citing buffy-progress file. Expected to clear the buffy queue within 1-2 days.

Peer remaining backlog — incorporated into master plan

Per peer reports/session_summary_2026-05-11.md:

P0 (immediate, blocking)

1. Confirm Block A INDEX→INDEX_STOCK fix didn't break peer's PR #904 — verify before next force-push window.
2. Verify drift_alert precondition runs in production — next hourly .github/workflows/audit-dashboard.yml regen check tier2_proven_strategies.cards[*].walkforward_gate.advisory_only.

P1 (this week)

1. Phase 1.5.1 — 7-consecutive-day drift_alert=false history check (needs daily snapshot scaffolding).
2. Phase 2.1 — extend alpha_engine/walkforward_validator.py to add COMMODITY + BOND + ETF to by_class output (currently 4 of 7 classes).
3. Phase 2.2 — implement walk_forward_by_strategy() (PROPOSED-NEW; does NOT exist yet). Pairs with peer's per-strategy edge_stability table.
4. Phase 4.1 — env-gate DRIFT_AUTO_PAUSE_DRY_RUN=1; 7d "would-pause" logging. Buffy E5 overlaps.
5. Phase 4.2 — class_capital_gate(asset_class) with capital_gate_log.jsonl. PROPOSED-NEW.

P2 (next 2 weeks)

1. Phase 5 Wave 1 (#1 + #2) — rolling-window profiling at 7/30/90/365/1095d + edge-decay heatmap.
2. Phase 5 Wave 2 (#3 + #5) — cross-symbol std-dev block + regime tag persistence on closed picks (closes regime_validation.regime_wr_breakdown all-zero rows).
3. Phase 5 Wave 3 (#7) — top-N portfolio Monte Carlo simulator (settles concentration debate before any real-money sizing — directly mitigates multi_asset_cot PF 19.19 / n=130 outlier risk).

Contingent (user-gated)

1. T1 — Flip DRIFT_AUTO_PAUSE_ENABLED=1 (HIGH risk; Phase 4.1 must be clean for 7d first).
2. T2 — Mutation autopsy on 14 quarantined strategies (LOW risk, read-only; due 2026-05-17, 7d post-quarantine).
3. T3 — Verify multi_asset_cot PF 19.19 via MySQL (LOW risk; needs DB_STOCKS_PASSWORD). Master plan P0 #5.

Areas worth further investigation (from peer summary)

• Walk-forward by_class drifting in real-time during regime collapse — argues for snapshotting at known cutoffs not live-reading.
• Cerebras rolling KS-D 30d monitor + xai stress-test layer — both queued for Phase 5; needs swarm round to choose before drift-pause flip.
• INDEX_STOCK class is empty (n=0) — remove or scaffolding? Investigate whether any signal generator emits asset_class="INDEX_STOCK".
• Shared .git peer-collision risk — 4+ silent branch auto-switches observed. Recommend isolated worktrees (git worktree add) protocol per agent.
• Mixed commit ownership — peer's git add -A swept other agents' files. Add project rule: explicit git add <path> only.
• dashboard_data.json local staleness — pre-Phase-2 protocol: git checkout origin/main -- audit_dashboard/data/dashboard_data.json to refresh just that file before any planning read.
• No backtest-cutoff disclosure on /audit — some systems show last_signal_at stale 2.5mo (e.g. ml_crypto_pred_v12) but cumulative WR/PF still surfaced. Misleading. Phase 4.3 DB-lineage card needs to include per-system cutoff date.
• Goal #2 (sports) + Goal #3 (events) untouched this session. Worth a parallel measurement-infra gap check before next session.

✓ Wave 1 SHIPPED — circuit-breaker HALT state removed

Commit 81bd0b86388 on main 2026-05-11 21:30Z. Removed alpha_engine/data/circuit_breaker_state.json (HALT, max_picks=0, 48d stale from 2026-03-24).

Verification PASSED: direct query against ejaguiar1_stocks::bt_backtest_trades shows MAX(imported_at) WHERE status IN ('WON','LOST') = 2026-05-11 20:00:59 with n=1,819,839. Forward validator unfrozen; WON/LOST writes resumed within 1 hour. Wave 1.5 independent fixes (lm_signals expire-cron, signal_tier writer, at_consensus_picks time-travel) still queued per kilo carveout.

Subsequent shipped commits

⚠ Kimi swarm audit (4-agent, 2026-05-11) — RAW-DB read contradicts dashboard verdicts

Archive: reports/kimi_edge_audit_2026-05-11/ (32 files — comprehensive_analysis_report.md, edge_audit_report.md, industry_standards_research.md, 7 rolling PNGs, 7 CSVs).

Kimi top-line claim: "4x inflated dashboard vs RAW DB"

Reconciliation: same DB, different filters

The gap is real but BOTH reads are technically correct:

• Kimi RAW reads include all 55,510 closed picks in at_raw_picks / trading_picks — including ghost rows, pre-resolver-v2 (Apr-27 fix) noise, dead-strategy artifacts, post-Wave-1 phantom EXPIRED rows. This is the pessimistic floor.
• Dashboard reads use asset_class_health (post-resolver-v2 filtered, CLAUDE.md verdict-grade) — Cursor reads n=7,875 CRYPTO vs Kimi n=51,049. The 7x difference is the resolver-v2 + ghost-quarantine filtering.
• Peer edge-stability sidecar reads further filtered (only currently-active sources, rolling window) — n=1,521 CRYPTO PF 1.39.

Master-plan adjustment: the dashboard "STABLE_EDGE" verdicts on COMMODITY + EQUITY still hold for the currently-active sleeve. But Kimi's WR 11.1% / Sharpe -2.34 RAW floor must be disclosed on /audit alongside the post-filter numbers — currently it isn't, which is the "4x inflation" Kimi flagged. Codex's readiness.by_class payload contract (P1) is the right place to surface both reads side-by-side.

Kimi findings that ARE net-new (not in current plan)

1. ML calibration system-wide inverted (not just ETF/CRYPTO per memory). 96% confidence → 0.9% actual WR. Promote calibration verification from P2 → P0 #9. Memory project_performance_reality confirmed it on ETF/CRYPTO subset — Kimi extends to all classes.
2. alpha_engine/anti_overfit_validator.py EXISTS but orphan. 13,886 bytes, last modified 2026-05-02. Contains CPCV / PBO / DSR code per Kimi inspection. P1 wire-up per CLAUDE.md Wire-Up Rule: production caller in calculate_smart_score / passes_smart_gate / score_pick. Once wired, automatically rejects strategies with PBO > 0.05 — the most critical defense against the 26-baby_strats-overfit pattern.
3. Empty model files: random_forest.py, xgboost_model.py, lstm_model.py are 14-byte placeholders per Kimi. P2 verification — confirm + remove or implement.
4. paper_trades table is EMPTY. Codex's SHADOW state requires 14-30d forward shadow tracking. Without paper_trades populated, SHADOW cannot be measured. P1 add — wire paper-trade variance recorder before any class can reach SHADOW.
5. Concentration risk: USDCHF=X = 215% of total PnL impact; INJUSDT = 89% of alpha_engine PnL. USDCHF=X almost certainly the FOREX unit-corruption bug (PR #876 pnl_pct clamp [-100, 200]%). P0 dependency — merge #876 before any FOREX verdict.
6. Day-of-week pattern (CRYPTO): Tuesday -3.37%, Friday -9.42%, Wednesday -13.85%, Monday -16.92%. Wednesday + Monday are worst. Future enhancement: time-of-day / day-of-week gate.
7. EQUITY t-test p=0.115 closest to significance (Sharpe +0.67) but RAW WR 1.84% (15 wins / 814 picks). Means the EQUITY POST-FILTER numbers (Cursor 54%, peer 57.4%) come from a small sliver of "good" strategies — the other 799 unfiltered picks are noise. Filter-criteria audit needed before EQUITY promotion. P0 #10.

Kimi DSR / PBO / WFE framework — adopt for real-money gate

Adopt Kimi's 10-step validation pipeline as the real-money gate before LIVE_ELIGIBLE: pre-register hypothesis → in-sample → WFA → CPCV → DSR → structural break tests → sensitivity analysis → transaction cost analysis → 3-6mo paper trading → graduated deployment (5% → 25% → 100%).

Net: Kimi's verdict is harsher than master plan but doesn't invalidate the COMMODITY/EQUITY STABLE_EDGE finding for the FILTERED sleeve. It does add 3 new P0 actions (calibration verify, EQUITY filter-criteria audit, PR #876 dependency surfaced) + 1 P1 (anti_overfit_validator wire-up) + a stricter real-money gate (10-step / DSR / PBO / WFE).

▶ /audit Decay Alerts — Action Required (2026-05-12)

14 HIGH rolling-7d WR drop alerts (>20pp baseline decay) + 3 MEDIUM staleness alerts surfaced on /audit. Per CLAUDE.md MUTATION_THREE_AXIS_PROTOCOL: mutation-before-kill applies; REDUCE-not-BLOCK is the soft action.

Recommended next-session work

• 2 P0 hard-blocks: futures_momentum (FUTURES) + MeanReversionBB (regime-driven decay) into BLOCKED_ASSET_STRATEGY_PAIRS. hs_lb_None investigation precedes block.
• 1 P1 soft-demote framework: add DECAY_ALERT_REDUCE set to audit_trail/quality_gates.py; deduct 8 pts from calculate_smart_score for any strategy in the set. Cleaner than hard-block since these may rehab once drift clears.
• 3 MED monitor: revisit at 2026-05-13 12:00Z (24h post-Wave-1) — silence likely resolves via outcome_resolver backlog flush.

▶ HIGH CONVICTION filter audit — 2026-05-12 vs session data

Current /audit HC panel shows per-class thresholds from 2026-04-15 data. Re-verified against session work (Kimi RAW DB read, anti-overfit DSR sidecar, P0 #10 EQUITY filter trace).

Summary — required HC filter updates

1. P0 most urgent: FOREX downgrade from "EDGE" to BLOCKED/DEAD. The current EDGE claim is the single most misleading verdict on /audit and could lead a user to size into a sub-floor (PF 0.28) class.
2. P1: ETF reclassify DEAD → THIN_REHAB (verdict was based on stale n=19 data).
3. P1: COMMODITY carve-out — admit cot_positioning family as EDGE while keeping class aggregate WEAK.
4. P1: CRYPTO + EQUITY + COMMODITY add anti-overfit DSR gate (require DSR ≥ 0.95 per anti_overfit_audit.json).
5. P2: Bump "2026-04-15" refresh date to "2026-05-12" once gates rewired.
6. P2: Refresh BOND n=8 → 18 (live).
7. P2: Display DSR verdict-counts inline (e.g. "CRYPTO: 4 EDGE_LIKELY_REAL / 33 OVERFIT_LIKELY") sourced from anti_overfit_audit.json::verdict_counts.

Implementation surface: audit_dashboard/hc_filter.js (per-asset-class HC gates per CLAUDE.md) + audit_dashboard/template.html:~1203-1300 (HC overlay text). Each verdict update is a small JSON-config tweak per config/hc_thresholds.json if it exists.

▶ DB Health red-tier — remediation status (refreshed 2026-05-12 03:30Z)

Per-metric progress on the 6 red-tier items in the DB Health — 2026-05-08T15:00Z panel. "Action required" commentary added inline. Updated 2026-05-12 03:30Z with sign-coherence guard, ghost-row triple-axis block, and CI commit-list fix.

Action required (master-plan commentary — refreshed 2026-05-12 03:30Z)

• Resolved (1): Forward Validator Freshness — full fix via Wave 1 unfreeze.
• Partial / forward-fixed (4): WON-vs-PnL (sign-coherence guard 22b677c1167), PnL Integrity (PR #876 + P0 #7), Phantom EXPIRED (PR #891), Ghost Rows (triple-axis quarantine 597819d79c7, 5 cohorts ~220k rows). Forward writes guarded; legacy rows still in DB until a backfill SQL pass.
• Diagnosed / implementation queued (1): Raw-Pick Outcome Coverage — root cause confirmed (missing upstream sync_active_mysql_picks_to_json()). Needs a new bridge that reads ACTIVE at_raw_picks, detects TP/SL/time-exit per asset class, and writes terminal entries into closed_picks.json + back into at_raw_picks. Promoted to P0 follow-up — the resolver itself works; it just has an empty queue forever.
• CI plumbing fix (also 2026-05-12 03:30Z): commit d317560ac9c adds audit_dashboard/data/db_health.json + audit_dashboard/data/cot_paper_pilot_status.json to the workflow commit-list (.github/workflows/audit-dashboard.yml:600). Both files were being FTP-deployed each cycle but never committed to git, so origin/main stayed 2026-05-08T15:00Z for 4 days. Next hourly cron will commit refreshed metrics — this DB Health panel will finally reflect post-fix state instead of the stale snapshot.

Recommended next checkpoint (post-fix re-snapshot)

Re-pull db_health.json at 2026-05-12 06:00Z after 2-3 hourly cron cycles. Expected deltas if fixes hold:

• Generated_at timestamp: 2026-05-08T15:00:00Z → current hour (CI commit-list fix landed)
• Forward Validator Freshness: 840h → <6h (RESOLVED)
• WON-vs-PnL contradiction: YES → likely still YES on first snapshot (legacy rows) but new resolver logs will show won_pnl_contradiction: WARNING count = 0 after one cycle if guard is hot
• Phantom EXPIRED: 100% → <80% (still high until Wave 1.5)
• PnL Integrity: 42% → 45-50% (forward writes only; legacy unchanged)
• Ghost Rows: 655k → ~440k (5 cohorts ~220k filtered from CRYPTO bucket)
• Raw-Pick Outcome Coverage: 0.09% → 0.09% (UNCHANGED until sync_active_mysql_picks_to_json ships)

⚠ DB Health red-tier crisis — original P0 remediation plan (Wave 0.5 → 4)

Live /audit DB Health panel (snapshot 2026-05-08T15Z but values persisted) shows 6/6 red metrics. This is the truth layer Codex says must be fixed first. Remediation already documented at reports/db_evidence_graded_final_2026-05-08.md (222 lines).

Wave 0.5 — Pre-deploy verification (5 min, READ-ONLY)

1. Confirm alpha_engine/data/circuit_breaker_state.json is git-committed (last touched commit fa9b6b38109 2026-03-28).
2. Run staleness SQL: SELECT MAX(imported_at) FROM bt_backtest_trades WHERE status IN ('WON','LOST') — expect 2026-04-02.
3. Gating-check grep: grep -n "circuit_breaker\|is_locked" alpha_engine/{forward_validator,outcome_resolver,production_scanner}.py.
4. Survivor signal_tier check: SELECT signal_tier, MIN(ts), MAX(ts) FROM at_discord_notifications WHERE signal_tier IS NOT NULL GROUP BY signal_tier.

Wave 1 — Unfreeze (5 min)

1. rm alpha_engine/data/circuit_breaker_state.json
2. Commit: fix(circuit-breaker): rm 2026-03-24 stale HALT state — unfreezes forward_validator (35d freeze)
3. Push to main; watch one cycle of .github/workflows/audit-dashboard.yml (hourly cron).
4. Verify: MAX(imported_at) for WON/LOST advances past 2026-05-11.

Precedent: PR #497 fixed R3 stale-state on 2026-04-27 (same class of bug). Same 2026-03-24 leak still live — referenced in freeze_2026_04_02_root_cause_2026-05-08.md:115. Memory ref: feedback_circuit_breaker_stale_state_leak.

Wave 1.5 — Independent pipeline checks (~30 min each)

Kilo carveout: lm_signals + at_consensus_picks + at_discord_notifications fail independently — NOT auto-fixed by circuit-breaker deletion. Each has its own cron.

1. lm_signals expire-cron: exit_price=0 in 96.2% of expire-cron rows (F10) — patch cron to skip expire-write when exit_price unset.
2. at_discord_notifications.signal_tier: 99.99% NULL (F8) — locate writer; backfill schema.
3. at_consensus_picks time-travel: 57.3% rows with future-dated entries (F4) — writer guard against future timestamps.

Wave 2-4 — Schema + route fixes

• STORED terminal_outcome column to eliminate WON-vs-PnL contradictions at read time.
• EQUITY route rewiring → challenge_200_trades (per remediation doc).
• Drop phantom row sets (after Wave 1+1.5 stable for 7d).
• Ghost-row cohort (655k rows, 18 cohorts) audit: identify the constant-pnl_pct writer; either repair or hard-drop.

Tie-in to SUPREME EDGE ENHANCEMENT

• Validates Codex truth-layer-first: all 6 red metrics are data-trust failures, not alpha failures. Cannot promote any class to OOS_READY until at least Wave 1 + 1.5 clear.
• Corrupts walk-forward.by_class: resolver stall → stale closed_picks → frozen algorithm_rolling_perf → walk-forward sees pre-2026-04-02 data only. Master plan already de-ranked walk-forward to advisory; this explains why.
• Existing PRs adjacent: #876 (pnl_pct clamp), #862 (DB query bank), #891 (mysql_sync closed_at fallback), #884 (category infer), #892 (safe_db_archive Hermes gate). Merge #876 + #891 + #884 + #892 before Wave 2 to lock in safety surface.
• Promotes DB Health to a master-plan P0: insert as P0 #0 (highest priority — precedes the existing 8-item P0 cluster).

Verbatim chatlog cross-check (peer docs/chatlog_verbatim_2026-05-11.md, commit 77f42fa5c3e)

Subagent scan of peer's 526-line verbatim chatlog identified 3 mandatory blockers + 1 class downgrade:

1. Phase reorder mandatory (swarm-unanimous): Phase 4 (measurement infrastructure) MUST precede Phase 2 (fast-track classes). Already captured in §"Peer plan v2" above; reaffirmed here. No per-class scaling until measurement gates exist.
2. Phase 1.5 drift-clearance gate: require drift_alert=false for 7 consecutive days before Phase 2 kickoff. Treat walk-forward as advisory-only while drift hot. Halt + recalibrate if drift persists > 30d.
3. BOND state downgrade: verbatim chatlog reads n=12 (not n=11 from Cursor or n=18 from Kimi). Either way well below n≥100 floor. Master plan BOND state changed from REHAB → BLOCKED above.
4. INDEX_STOCK class rename adopted: 8 classes now (CRYPTO/EQUITY/COMMODITY/FOREX/ETF/BOND/FUTURES/INDEX_STOCK). Peer Block A commit 5e4bc1efe63 ships the rename + defense-in-depth.

COMMODITY / EQUITY — STABLE_EDGE but NOT auto-promoted to OOS_READY

Verbatim chatlog edge-stability verdicts (COMMODITY STABLE_EDGE n=167, EQUITY STABLE_EDGE n=272) do NOT auto-promote these classes to OOS_READY. Per Codex truth-layer-first policy + master-plan promotion gates, these classes hold at REHAB until:

• COMMODITY: multi_asset_cot PF=19.19 DB-verified + CT=F/KC=F concentration disclosed + walk-forward.by_class emits real folds
• EQUITY: claude_gainer_st winner-vs-blacklist contradiction reconciled + capped-vs-raw PnL gap verified + bottom-symbol pruning lands without OOS consistency < 80%

edge-stability sidecar verdict alone is necessary but NOT sufficient for promotion.

User policy quotes (from verbatim chatlog)

• Commit ownership protocol (line 211 of verbatim chatlog): "wait so daily_ideas.md is being written by the other agent, check in with them before changing that" → establishes coordinated commit ownership, not parallel overwrites. Adds to peer summary's recommendation: explicit git add <path> only; check with file's primary author before edits.
• Cadence directive (line 18): "proceed. create a set of todos based on this new items, and fire whatever is needed and continue going! ensuring progress every 30 minutes... talk to peers also" → affirms multi-agent coordination pattern.

Buffy enhancements — review confirmation

Full-file re-read of both buffy artifacts (chatlog 43 lines + progress 134 lines). Confirmed:

• 4 enhancements (E1/E2/E4/E5) all PENDING — Swarm Cross-Check log, Typecheck Results, Code Review Results, and Git Commits tracking tables are all empty (one row each, all "—"). Work has not yet started post-design.
• 3 issues directly verified by buffy code review (review file lines 33-38):
  • FOREX benchmark missing — benchmark_return('FOREX') returns None (no FOREX in benchmark_map at live_market_fetcher.py:155-162) despite DXY data fetched as "DX-Y.NYB"
  • Per-pick disk I/O — _drift_pause_active() at quality_gates.py:4143 reads dashboard_data.json on every passes_active_gate() call (synchronous file read in hot path; 60+ active picks = 60+ unnecessary disk hits/cycle)
  • No staging mode for drift pause — only on/off toggle; no dry-run to test before flipping DRIFT_AUTO_PAUSE_ENABLED=1
• Master-plan ranking unchanged: all 4 buffy enhancements stay P1. E1+E2+E5 directly support the master-plan FOREX P0 cluster and Codex truth-layer; E4 (excess-return alert < -5%) complements edge-stability sidecar as an early-warning monitor.
• No new buffy items beyond E1/E2/E4/E5. Tracking-table rows for E3 absent (skipped in numbering — likely an earlier idea dropped during scoping).

Why this plan is not generic (verification checklist)

• ✓ Each of 8 asset classes has its own action list with named strategies + specific file:line references + per-class promotion gate
• ✓ COMMODITY actions cite cot_positioning_CT_locked + multi_asset_cot + CT=F/KC=F concentration disclosure
• ✓ EQUITY actions cite rs-breakout-scout + aggregated_picks + claude_gainer_st contradiction + Breakout Momentum
• ✓ CRYPTO actions name kimi_signal_tracking + baby_strats:crypto_soc_* + quan_engine 18% drag + st_fear_greed_contrarian 94% WR + alpha_engine_fast drag #1
• ✓ FOREX actions cite PR #876 unit-corruption fix + memory feedback_noncrypto_resolver_live_close_bug root cause + CLAUDE.md mutate-before-kill directive
• ✓ ETF actions name specific symbols (XLF, XLE, XLK) + leveraged-ETF exclusion
• ✓ BOND + FUTURES + INDEX have explicit "keep paper / re-emission plan / defer" verdicts not just "expand sample"
• ✓ Cross-cutting P0 cluster names 8 specific actions with effort estimates + source-plan attribution
• ✓ Open PR triage assigns master-plan priority to all 24 open PRs, with recommended merge order anchored to per-class fixes
• ✓ Each promotion gate has measurable numerical thresholds (PF / WR / MDD / n / consistency / variance)