Strategy Playbook — Polymarket Earnings Trading Bot

Status: Single forward-looking source of truth for strategy improvements as of Aug 21, 2026 (PKG-357 #366 DEPLOYED — $10k capital, S2c sizing, B-4 table, C-1 re-close, RES M1+M5, CLB-1, B-5 pilot; see the Aug-17–21 banner below). Baseline before that: Jul 5, 2026 (Finding #330 — RuleB FAILED-GAUNTLET reversal). Prior baseline: May 31, 2026 (Findings #314.c sizing-sweep [KEEP vs lowering; RAISING gauntlet-PASS at live 1.40 cap] + #314.d hybrid-gate KEEP $1k + #314.e 4 SKIP-rule candidates analysis-only + #315 exposure-cap re-validation [KEEP 1.40, #313 cap-raise inert]; Findings #311 + #312 + #313 SHIPPED — gate $1k + v2 #6 + PR 3.a priority/capital).

🛑 Jul 5, 2026 — Finding #330: RuleB (±90 min no-stop window) FAILED-GAUNTLET — Phase-2 active flip CANCELLED (simulator gap-fill bug; NO live trading change). The operator-directed Jul 5 strategy review found simulate_window_only filled window-blocked stops at the STOP PRICE even when the price was still through the stop at rearm — a fill the as-shipped #305 cancel-and-rearm mechanism cannot achieve (a re-placed STP-LMT triggers instantly and fills near market; simulate_finding_a_hard models exactly this as rearm_gap_open, the arm that retracted FA_hard in #228 at −$1,545/100). Gap-aware fix + prod re-run (n=129): RuleB headline −$209/100 (was +$224), ex-top-2 −$194/100 (was +$122), ex-repeat −$227/100 — FAIL on every §K check; optimism bias ≈ $535/129 trades. RuleG also flips negative (−$30/100 headline). Live corroboration: all four June in-window stops (LULU/ADBE/ACN/KR) were correct stops — RuleB-active would have worsened each (ACN −$26 actual vs ~−$134 held). NO stop-loss rule variant currently survives the gauntlet. Consequences: RuleB Action-Queue row → FAILED-GAUNTLET; v3 D + v3 E lose their RuleB dependency (re-evaluate as pure selection rules at trade #150+ / SHORT n≥30); #323.b's "0.92-0.99 conditional on RuleB active" lever is dead. Phase-1 dry-run code remained (logging-only; duplicate-log spam fixed — 8,026 lines → once per position) until Aug 21, 2026 (#380): the whole RuleB apparatus was REMOVED (operator A-12 option (b); the harness had been structurally blind since mid-June — its logs validated nothing). See docs/FINDINGS.md #330 + #380. Sections below written before Jul 5 that call RuleB "the only surviving stop-loss rule" are historical narrative.

🔬 May 31, 2026 — Finding #319: PR 5 depth-gate v2 widened reconstruction — the "unreliable orderbook" skip is a roughly break-even low-information-signal filter → KEEP (offline-only; closes Finding #283; Phase 3 simulator PR 1-5 COMPLETE). Answers "has the depth/spread gate ever saved us from a bad trade?" at proper sample size. Depth-logging only began 2026-04-14 (~4 kills); widened by replaying the REAL ConvictionCalculator.get_conviction_at_entry_time point-in-time over the full 915-market universe. NEW scripts/phase3_simulator/depth_gate_variants.py + replay_depth_gate_v2.py. 22 binding kills: 13 losers BLOCKED (−$142.56), 9 winners KILLED (+$165.09), net removal +$22.54 over ~7 months (≈ noise). So YES it has saved us — reversing the n=4 read (which showed 0 losers blocked = small-sample mirage). Kill cause 18/22 snapshot_reliability (hollow 1¢/99¢ book → sparse/unstable price-snapshot conviction), 1 spread → a low-information-signal filter, not the pure spread gate #283 implied. remove_all +$22.54 / carveout_real_trade +$22.78 (≈ identical) → adds variance for ~$0 EV. Operator-confirmed KEEP; production gate UNCHANGED. 395 tests PASS. See docs/FINDINGS.md #319.

🔬 May 31, 2026 — Finding #318: PR 4 regime classifier v2 sweep — faster detectors add no robust ROI → KEEP compound_v1 (offline-only; closes Finding #278). Tests whether a faster downtrend detector improves Rule #1 (Skip LONGs in downtrend, #304) vs the production compound rule (median 31-trading-day lag from SPY peak). NEW scripts/phase3_simulator/regime_classifiers.py (point-in-time SpyDailyHistory + 6 classifiers) + replay_regime_v2.py (re-classify point-in-time prior-close → re-apply Rule #1 → recompute P&L via run_sizing_sweep/apply_sizing_gauntlet vs compound_v1, on live-only primary + capacity-aware 1.40 cross-check). Prod (568 signals): below_ema20 (fastest; skips ~40% of LONGs vs compound's 17%) gains +$70.31 PASS on live-only but −$17.02 FAIL-EX-TOP-2 on the capacity-aware universe — the live-only win is selection luck that doesn't generalize (same reversal as #315 "don't look in isolation"). compound_or_dd5 (#278 option B) + below_ma50 inert; dd_3pct + compound_symmetric FAIL. The only durable effect is a better worst-week tail (5p −$98.60→−$57.08) at ~breakeven P&L (downside trade-off, not a money-maker). Methodology validated: PIT compound_v1 reproduces the live recorded regime 95.3%. Operator-confirmed KEEP compound_v1 (Rule #1 optimizes for ROI; faster detection adds no robust ROI). Closes #278; VIX deferred to Phase 4. 375 tests PASS. See docs/FINDINGS.md #318.

🔬 May 31, 2026 — Finding #314.e: 4 deferred SKIP-rule candidates (PR 3.b.3) — ANALYSIS-ONLY, ship deferred to operator. Re-evaluated the 4 Finding #313 deferred SKIP cells REAL-SCALED (not hypothetical $10k-notional) + capacity-aware + 5-check gauntlet. Real-scaling corrected 2 cells: Large+SHORT+$2k-$5k is only n=3 (not 12 → can't gauntlet); Large+LONG[0.80,0.90)+$1k-$2k is net-POSITIVE on real data (+$0.93 avg; hyp said −$20 → skipping it HURTS, FAIL-EX-TOP-2). Gauntlet-PASS ship-eligible candidates: skip Large+LONG[0.80,0.90)+vol≥$10k (+$2.04/wk, PASS, improves maxDD); skip Mid+LONG[0.80,0.90)+$1k-$2k (+$1.12/wk, PASS); combined skip-all-4 (+$5.42/wk, PASS, best maxDD −$220, WR 39.1%). Per operator ship-bar (C) nothing shipped — per-candidate ship decision pending operator AskQuestion. See docs/FINDINGS.md #314.e. (Partially advances the sizing-priority-v2 deferred item — the per-direction/per-cell SKIP-rule arm.)

🔬 May 31, 2026 — Finding #314.d: Hybrid seasonal gate variant (PR 3.b.5) — KEEP uniform $1k (NO production change). Operator idea: gate $2k in peak earnings-weeks / $1k off-peak (peak periods have plenty of $2k+ markets so $1k "just adds noise"). Verified the academic peak-week classifier tracks PM data first (peak windows = 57.7% of 977 markets / 60% of vol≥$1k). On the 568-signal simulator universe the hybrid removes 144 peak-week $1k-$2k signals → +$8.08/wk vs uniform $1k +$16.86/wk (Δ −$8.78/wk, FAIL-SPILLOVER), nearly reverting to uniform $2k (+$6.70/wk). Premise REFUTED — those peak-week $1k-$2k markets are net-positive (#314.b $1k-$2k bucket +$423), so removing them gives back most of Finding #311's lift; the capacity-reallocation hope doesn't materialize (cap-skips 58→5 but freed slots not refilled, maxDD worsens). Decision: KEEP uniform $1k (Finding #311 unchanged). See docs/FINDINGS.md #314.d.

🔬 May 31, 2026 — Finding #314.c: Sizing variants sweep (PR 3.b.4) — KEEP-30/22/15 (the then-live tiers, since superseded by flat-25 #314.f → S2c #366) holds only vs LOWERING; RAISING is gauntlet-PASS at the LIVE cap (NO production change). 🔁 CORRECTION (per Finding #315): the original run used exposure_cap=2.0 but the live cap is MAX_DAILY_EXPOSURE=1.40. Re-run at 1.40: lowering sizing still FAILS (all FAIL-SPILLOVER), but flat 25/25/25 (+$4.18/wk) and 35/25/18 (+$4.63/wk) now PASS the gauntlet — at the tighter real cap, concentrating capital into bigger positions wins. These are now ship-eligible candidates (operator decision; caveats: exceed the stale MAX_POSITION_SIZE_PCT=0.15 + small-cap 5% stop/squeeze risk). First sweep on the size-aware capital simulator (offline). Operator hypothesis: capacity binds at 99% (Finding #314.b) → lower sizing to fit more concurrent positions and lift P&L. REFUTED. On the 568-signal / 36-wk universe ($2.5k × 2.0 = $5k max gross), lowering sizing admits more positions but total P&L drops — 20/15/10 fits +28 more trades (437 vs 409) yet loses −$6.80/wk; all four lower-sizing variants (25/18/12, 20/15/10, 22/16/11, flat 18/18/18) FAIL-SPILLOVER (marginal capacity-freed trades are low-EV per #314.b, shrinking good positions costs more). Raise/flat variants give only tiny non-robust lift (flat 25/25/25 +$1.41/wk FAIL-EX-REPEAT-TICKERS; 35/25/18 +$0.75/wk WARN-OUTLIER-DOMINATED + worse maxDD/5p). All 8 sit on the Pareto frontier (lower sizing trades E[wk] for better DD/5p — so the bias gauntlet, not Pareto, is the decider; no variant has a gauntlet-PASS positive lift). Decision: KEEP Large 30% / Mid 22% / Small 15% (Finding #211 unchanged). Partially closes the sizing-priority-v2 deferred item (sizing arm). Simulator gained scale_pnl_with_size (design §12 PR3) + gauntlet.apply_sizing_gauntlet; 296/296 tests PASS. See docs/FINDINGS.md #314.c.

🚀 May 30, 2026 LATE LATE evening — Finding #313 SHIPPED: PR 3.a — Volume-tier priority bump + capital cap raise + NYSE dead-code cleanup. THIRD real-money strategy ship of the day. Priority scale 0-6 → 0-8 (kept Mar 9 mcap × conviction ordering since live LONG WR Large 45.7% > Mid 34.3% still holds; ADDED volume bump since live LONG WR jumps 36% NULL → 64% $2-5k → 83% $5k+). MAX_TOTAL_EXPOSURE_PCT 1.40 → 2.00 (matches Reg T overnight ceiling, +43% capacity). NYSE/NASDAQ branching dropped (verified dead post Finding #185 Feb 15 unification). 617/618 tests PASS. Operator follow-up: re-validate $1k gate (#311) against new volume-tier data — $1k-$2k bucket has been the ONLY losing bucket historically. 4 SKIP rule candidates DEFERRED until simulator framework is ready.

🚀 May 30, 2026 LATE evening — Finding #312 SHIPPED: v2 #6 Daily 3-stop session breaker + DirectionPolicySpec doc-bug fix. Second real-money strategy ship of the day. Pure-decision module src/earnings/session_stop_breaker.py + _count_todays_stops() helper + skip block in signal_generator._build_trading_signal after Queue Rule #1 long_downtrend. Defaults SESSION_STOP_BREAKER_ENABLED=True, SESSION_STOP_BREAKER_THRESHOLD=3. Historical evidence: 9 sessions in production with ≥3 stop-exits in a single ET-day (2026-02-03 through 2026-04-29). Estimated +$1/wk + ~-$50 Max DD improvement (heuristic). Plus DirectionPolicySpec correction: production_strategy_may30() factory now accurately models prod thresholds (0.70/0.30 with 0.90 over-confident skip + raw_prob conviction_fn) instead of the wrong Phase α defaults (0.55/0.45/abs_distance_x2). 79 tests PASS local (38 phase3_simulator + 22 session_stop_breaker pure + 8 integration + 11 regression). NO integrity-monitor rule count change.

🚀 May 30, 2026 evening — Finding #311 SHIPPED: gate_uniform_$1000 first real-money strategy ship since Mar 9, 2026. SIGNAL_MIN_VOLUME_TOTAL default lowered $2,000 → $1,000 in src/earnings/config.py. After full operator-driven correctness audit (6 errors found, all fixed) + Phase ε/α/β/γ/δ/ζ unified strategy framework completion. Per-tier empirical scaling (Large $537 / Mid $339 / Small $205) replaced wrong hardcoded $245. Phase γ+δ replay shows: Δ E[wk P&L] +$21/wk (~+$1,092/yr at $2,450 capital), Δ max DD -$46 (within -$735 = 30% capital target), outlier alarm fixed (production ex-top-5 -$152 → +$582 post-admit), Pareto-optimal among 12 candidates. Other 2 ship-eligible variants RETRACTED-IN-FAVOR-OF gate_uniform_$1000 (operator picked simplest single-axis change). 38/38 phase3_simulator tests PASS. Production strategy now lives in production_strategy_may30() factory; production_strategy_may29() is backward-compat alias.

Verified still current as of May 30, 2026 — Phase 3 simulator PR 2 COMPLETE (Finding #310.c). PR 1 (foundation + cross-check + gauntlet) shipped May 28; PR 2 (volume-gate retune sweep) shipped May 29 across 3 sub-commits (PR 2.a gate_variants + PR 2.b counterfactual_admission + PR 2.c sweep + dashboard). The seasonality preflight that gated PR 2 (operator-requested INFERENCE DISCIPLINE check) REFUTED the pure-seasonality hypothesis — Dec 2025 off-season was healthy at 98.1% gate-pass + $14,522 median volume, while Mar+2026 off-season collapsed to 27.8% + $0 median. This is a REAL STRUCTURAL collapse starting March 2026, NOT seasonality. The PR 2 sweep (262 usable counterfactual markets, all in collapse period; pre-coding aggregate +$3,180 P&L if all admitted) produces ship-recommendation candidates per the 3-condition conservative-ship policy (Q3 (C): full-pop gauntlet PASS + ≥20 admitted in collapse + ex-top-2 stays positive) — operator AskQuestion required per qualifying variant before any production gate change. Dashboard 🧪 Phase 3 — Gate Retune Sweep section is live. PR 3-5 now COMPLETE (#306-#319) + Phase 4 dashboard consolidation COMPLETE (#306-#322) — PR 3.a priority/capital (#313), sizing/exposure-cap evaluation (#314.c/#315/#317), PR 4 regime KEEP compound_v1 (#318), PR 5 depth-gate KEEP (#319). See §1 banner above + docs/PHASE_3_SIMULATOR_DESIGN.md §12.

Earlier — May 28, 2026 LATE evening — Phase 3 simulator PR 1 COMPLETE. No strategy rule added or retracted between Apr 25 and May 28; two queue rules SHIPPED today (Rule #1 + RuleB Phase 1 DRY-RUN) and the Phase 3 unified simulator PR 1 foundation landed across 5 commits (#306.a foundation skeleton + #306.b cost-model + replay + reference_adapter + cross-check script + #306.c gauntlet + H-rule self-validation + reproducibility + JSON output + #307 #273 conditional-label fix + #308 Bug #6 hyp_pnl re-run + structural-ceiling discovery). All 4 PR 1 acceptance gates PASS: §10.1 cross-check vs analyze_finding_a_full.py (100% match within $1 on all 7 rules / max drift $0.00 across 121 trades on prod May 28 LATE evening); §10.2 reproducibility (2 runs = 0-byte diff after masking generated_at — verified on real prod data + unit-test); §10.3 H-rule self-validation (FAIL-EX-TOP-2 per test_h_rule_synthetic_returns_fail_ex_top_2); Phase 0 §0.6 Gate 1/4 corrected per #308 structural ceiling. Live baseline gauntlet on prod: RuleB verdict PASS, per-100 $+90.74 headline, $+39.93 ex-top-2 vs A_base baseline — ship recommendation SHIP. (The reference simulator's narrative numbers were variant-vs-LIVE: +$236/100 / +$127/100 ex-top-2; the gap is the intentional difference between replay-vs-baseline and replay-vs-recorded semantics — both are correct; replay-vs-baseline is the right default for PR 2's gate-variant comparisons.) PR 2 (volume-gate retune sweep — highest-leverage per May-28 PM volume-collapse evidence) begins next session per docs/PHASE_3_SIMULATOR_DESIGN.md §12.

Apr 26 → May 28 intervening work, none of which was a strategy rule change until the May-28 Phase 3 sprint: - Strategy Analysis Revamp Phase 1/2/2.x — Findings #234-#247 - Database Integrity Audit Phases 0-6 — Findings #250-#261 - Dashboard 500 incident chain #262 - Rigorous §B re-audit + ancillary fixes — Findings #263-#274 - Apr 28 deferral parking — Findings #278 (regime detection v2) + #279 (sizing & priority v2) - Apr 29 dashboard UX — Findings #280 (outcome badges) + #281 + .v2 + .v3 (trade-chart modal) - Apr 30 volume policy fix — Findings #282 + .v2 (max(DB, Gamma) + shared single-shot fetcher) + UX rename #283 ("LOW OB" → "WIDE SPREAD") - May 6 P&L Summary enrichment — Finding #284 (tooltips + Trading Capital ROI panel) - May 9 exposure-cap correctness — Finding #285 Phase A (greedy-fill in BMO + AMC loops + opportunity_log emission) - May 9-10 — Pre-Phase-3 housekeeping batch + Dome→Polymarket migration + plan-(C) follow-ups + new follow-ups all shipped (Findings #269 / #270 / #273 / #274 / #286 / #287 / #288 / #288.v2 / #288.v3 / #289 / #289.v2 / #289.a / #289.b / #290 / #291 / #291.v2 / #292 / #293 / #294 / #295). NONE of these are strategy rule changes — all infrastructure / data-quality / cleanup / observability. - May 11-28 — Operational hygiene + observability close-outs (5 findings + 1 doc-only rule codification + 2 factual corrections): #296 (May 11 — _check_systemd_unit_last_result branch reorder); #297 (May 17 — api_health_checker WARN-but-exit-0 cry-wolf fix); May 27 factual correction commit d3e3e9c (Sunday 2FA reconnects DO require operator push approval — corrected the May-17 / May-24 entries); May 28 codification of the 🚨 INFERENCE DISCIPLINE rule in .cursor/memory.md (commit 875f6e3, companion to NO SHORTCUTS, triggered by an agent false-alarm on AMC scheduling that should have been a 60-second model-vs-code verification); #302 (May 28 — polymarket-backup.timer collision with apt-daily-upgrade.timer, moved to 05:00 UTC); #303 (May 28 — deploy_from_github.sh pre-pull hygiene for recurring working-tree drift). Plus 2 IBKR Sunday 2FA reconnects (May 17, May 24) and 1 mid-week auto-recovery (May 28 04:12 UTC, 6m 17s downtime, no operator action). NONE are strategy rule changes. - May 28 — Phase 3 sprint (5 PR-1 commits + Phase 3 unified simulator PR 1 COMPLETE): Finding #279 Phase B amendment (PM earnings-market volume collapse evidence folded in); Finding #304 Rule #1 Skip-LONGs-in-downtrend SHIPPED active (first Phase-3 strategy rule); Finding #305 RuleB ±90min no-stop window Phase 1 DRY-RUN SHIPPED (second Phase-3 strategy rule; hardware cancel-and-rearm pattern; RULEB_ENABLED=false in prod by default); Finding #306.a Phase 3 simulator foundation skeleton (scripts/phase3_simulator/ package + RuleSet dataclasses + CLI stub); Finding #306.b cost_model + replay + reference_adapter + cross_check_baseline script (§10.1 acceptance gate satisfied: 100% match within $1 on all 7 rules vs analyze_finding_a_full.py across 121 trades on prod); Finding #306.c gauntlet.py (5 checks per memory §K) + output.py (deterministic JSON schema) + runner.py --run-gauntlet (§10.2 reproducibility 0-byte diff + §10.3 H-rule FAIL-EX-TOP-2 acceptance gates satisfied; live RuleB verdict PASS on prod); Finding #307 snapshot_schema conditional-label fix at signal_generator.py:1041 (closes original Finding #273 Phase 1); Finding #308 Bug #6 hyp_pnl re-run + structural-ceiling discovery (~69% generate_signal / ~28% depth are deadband-conviction ceiling, not gaps). All 4 PR-1 acceptance gates PASS; PR 2 (volume-gate retune sweep) begins next session.

The Action Queue in §2 has 2 rules moved from IN-QUEUE → IMPLEMENTED today (Rule #1 + RuleB Phase 1 DRY-RUN). May 28 regen of scripts/strategy_performance_analysis.py + scripts/analyze_finding_a_full.py already done — fresh numbers in the §1 verdict-banner above; RuleB ex-top-2 strengthened to +$127/100 (gauntlet PASS); ~117 closed trades / +$212 net / ~37 % WR / 51.7 % stop rate / pre-trade-#150.

✅ UPDATE (May 31, 2026) — this deferred bundle is now CLOSED: $2k volume-gate retune SHIPPED $1k (#311); 0-6 tiers re-validated + extended to 0-8 (#313); rare/normal/busy exposure cap collapsed to uniform 140% (#317, after #316 showed the rare discount near-inert); regime detection v2 → KEEP compound_v1 (#318); depth-quality gate v2 → KEEP (#319). The historical parking narrative below is preserved verbatim.

Apr 28 → May 28, 2026 — four larger items parked for post-Phase-3 review (§7 "DEFERRED-POST-PHASE-3"): regime detection v2 (current classifier lags SPY peaks by median 31 trading days; Finding #278), sizing + priority tier system + exposure-cap re-evaluation v2 + $2k volume-gate retune (Mar 9 percentages + 0-6 tiers not re-validated, AND rare-day exposure cap fires on majority of trading days per Finding #285 Phase B, AND PM earnings-market volume collapsed ~109× Jan→May with $2k-gate pass rate now 1.9 % — May 28 evidence folded into Phase B; bundled together since they all gate which valid-signal trades actually enter; Findings #279 + #285 Phase B + PM-volume- collapse amendment), Polymarket depth-quality gate v2 (37/38 historical kills were spread-dominated; gate decomposition + threshold re-tuning needed; Finding #283), and pm_v2 schema NULL bid/spread on signal_emitted rows fix in lockstep with the simulator ship (Finding #273). All four surface on the dashboard 🎯 Action Queue under the DEFERRED-POST-PHASE-3 status filter.

Combines: v3 audit (docs/STRATEGY_AUDIT_v3.md, Apr 22) + Apr 25 Finding-A retraction simulator (docs/FINDINGS.md Finding #228) + Apr 25 evening queue re-validation (Finding #232) + corrected stop-loss classification (Finding #233) + the dashboard Strategy Analysis tab. The dashboard tab is the operational view of this same data; this file is the narrative + decision rationale.

Audience: - The next agent / human picking up the project. - The operator deciding what to ship next.

Read order: 1. §1 current verdict 2. §2 ranked action queue (= what to ship next) 3. §3 verified findings — mechanism + evidence + bias check 4. §4 retracted / failed findings — so they don't get re-suggested 5. §5 mandatory bias gauntlet 6. §6 implementation notes per rule (file + LOC + acceptance test) 7. §7 deferred questions / data gaps 8. §8 cross-references (scripts, docs, dashboard)

0. 🚨 NO SHORTCUTS — EVER (the top rule, applies to strategy work too)

Before any strategy edit on this doc or any rule change in build_action_queue(): no shortcuts, ever. If the operator says "verify all 4 bias-gauntlet checks," verify all 4 — not "the top 2 are clearly fine." If they say "re-run the simulator on every variant," re-run every variant — not "the headline ones." If they say "ex-top-2 numbers alongside headline," include both — not "I'll add ex-top-2 in a follow-up." Strategy work has the highest capital risk on this project; shortcut discipline matters MOST here.

Anchoring example + full rule: see .cursor/memory.md "🚨🚨🚨 NO SHORTCUTS — EVER 🚨🚨🚨" section (codified May 10, 2026 from a doc-cleanup shortcut incident — the operator's framing was "if there is one thing that you must always know about this project is this: no shortcuts! ever!").

1. Current verdict (Apr 25, 2026 — late evening; post Findings #229–#233; May 28, 2026 refresh on same trade population — see banner below)

May 28, 2026 refresh (live regen of scripts/strategy_analysis_report.json and scripts/finding_a_full.csv): Strategy population on the same trade_audit_log-with-reconciliation-filter is now 116 clean closed trades (43 W / 73 L) at WR 37.07 %, net +$212.21, avg +$1.83/trade, with the all-trades (incl. 3 bug-affected) population at 119 / +$207.72 / WR 36.97 %. Stop rate 51.72 % (was 49 % Apr-25 baseline — slightly worse). The strategy has accumulated 0 new clean trades in the 33 days Apr-25 → May-28: last clean closure was 2026-05-07 (id=162 MCD), next entry was 2026-05-27 (id=163 HPQ, still open as of May-28 12:30 UTC). The 21-day entry drought is consequential, not idle — see the PM earnings-market volume-collapse evidence in docs/FINDINGS.md Finding #279 "Phase B framing premise" subsection (Jan→May avg cumulative volume dropped ~109×; May $2k-gate pass rate 1.9 % = 3/154 markets). RuleB simulator re-run on fresh data is now n=121 / per-100=+$236 headline / per-100=+$127 ex-top-2 (was n=51 / +$123/100 on Apr-25 baseline — RuleB has STRENGTHENED on robustness). RuleG headline +$242 ex-top-2 +$105 — higher headline than RuleB but less robust; RuleB was the recommended next stop-loss ship and has since SHIPPED as Phase-1 DRY-RUN (#305, May 28). FA_hard retraction reasoning still holds: rearm_gap_open arm n=24 at −$1,545/100. (May 31 update: live population is now ~117 closed trades — see live dashboard — and all 4 deferred-bundle items are now CLOSED, not DEFERRED: #311 gate $1k / #313 priority 0-8 / #317 uniform 140% / #318 regime KEEP / #319 depth KEEP.) The Apr-25 narrative below is preserved verbatim because the underlying conclusions (5-outlier curse, stops-eat-winners, selection-rules-first) are unchanged.

Strategy state on ~117 closed trades (May-28 refresh): net P&L +$212 at a ~37 % win rate (Apr-25 figure was +$167 / 33.6 %; morning handoff was +$61 / 109). The strategy picks direction correctly 83 % of the time — the leak is in conversion to P&L on the stop-loss leg.

Stops are the single biggest leak: across ~117 closed trades stops are the dominant P&L drag (current stop rate per the live dashboard / FINDINGS; corrected via exit_reason ILIKE '%stop%' — the stop_triggered boolean undercounted at 1.7 % and is broken; see Finding #233). Stops have cost the strategy −$676 in aggregate (avg −$11.86 per stopped trade). Non-stops have contributed +$843, so stops are erasing ~80 % of the gross alpha. Mar 9 #202 wider stops dropped frequency from 50 % → 45.5 % but raised average loss per stop from −$9.66 → −$22.23.

5-outlier curse (renamed Apr 25 from 3-outlier — Finding #232): GPRO + DELL + SWBI + TXN (NEW Apr 22, +$92, LONG large) + STUB (NEW Mar 4, +$58, SHORT mid downtrend) together carry ~$430 of the ~+$212 net. Ex-them the strategy is −$263 across 111 trades. Any analysis that includes them looks great; any analysis that excludes them is structurally negative. Always report ex-top-2 alongside the headline; for any slice containing 2+ of the named 5 outliers, also report ex-them.

Universe view state (Apr 25 late evening, post-backfill + organic growth)

After Findings #229 (volume backfill, +319 markets) and #230 (conviction backfill, +384 markets) and a few hours of organic resolution, the Universe View now covers 604 of 900 earnings markets (67 %, up from 31 % this morning). In-band markets pass the bot's $1k volume gate (#311; gate lowered from $2k → $1k May 30) — see the live Universe View for the current pass % (the Apr-25 snapshot was 90.4 %, 546 of 604, measured at the then-$2k gate), up from 35 % positive-volume coverage that morning. This is the operational signal-quality denominator for every "is the bot's signal worth pursuing?" question. Detailed breakdowns + live numbers refresh every report run on the dashboard 🌐 Universe View section.

What the most recent work proved (updated late Apr 25)

• The stop-loss leg is not fixable by tweaking stop semantics broadly. Apr 25 simulator (scripts/analyze_finding_a_full.py) showed v3 Finding A is either inert (rule as written) or net-negative (only effective form: −$418/100), and Findings D / G fail the 4-check bias gauntlet.
• RuleB (±90 min stop window around earnings release) is the only stop- loss intervention that survives the gauntlet, at +$123/100 trades — PROMOTED to rank 2 in the Apr 25 evening re-validation (was rank 4) because the corrected stop classification (Finding #233) revealed live stop bleed is the strategy's biggest leak (−$676 across 57 stopped trades). (SHIPPED May 28, 2026 as Phase-1 DRY-RUN — Finding #305; no longer queued — RULEB_ENABLED=false in prod until operator flips active.)
• v3 F (BMO + VIX-bucket sizing) was DROPPED to RETRACTED on Apr 25 evening (Finding #232 deep-dive bias gauntlet): the 1.2× scale-UP arm fails ex-top-3 (the +$100 winner bucket collapses to −$1.58 across 17 trades after dropping STT/VZ/ALLY); direct multiplier-on-realized-pnl modeling gives only +$36/100 (not +$60); operator dropped both arms.
• v2 #6 (3-stop session breaker) has REAL historical evidence after the corrected stop classification: 9 sessions in history with ≥3 stops (Feb 19 had 5). Was previously believed to be tail-hedge only with no historical fire. (SHIPPED May 30, 2026 — Finding #312; threshold 3, ACTIVE.)
• The path to improving P&L is selection rules + RuleB — cut bad-EV trades upstream (downtrend LONGs — now ACTIVE via Rule #1, Finding #304, shipped May 28 — then Mid-cap-LONGs / non-downtrend SHORTs once RuleB has validated whether their issue is selection or stops). Each selection rule has cleaner direct evidence than the stop-loss family.
• The conviction signal IS profitable at universe scale (post Finding #230 backfill): 529 LONG > 60 % markets total +$26,506 at $10K notional (avg +$50/trade); 39 SHORT < 30 % markets total +$5,020 (avg +$129/trade). Bot's modest +$1.44/trade actual is the gap between this universe signal and live execution (capital constraints + stops + selection).

Combined realistic improvement (revised Apr 25 evening)

The (formerly) IN-QUEUE rules (B + RuleB + 3-stop breaker) are now all IMPLEMENTED — Rule #1 skip-LONG-downtrend (#304), RuleB ±90min dry-run (#305), daily 3-stop session breaker (#312); F dropped — plus indirect RESOLVED-BY-OTHER-RULE potential on D + E if RuleB closes their stop bleed: estimated +$190 to +$310 / 100 trades realistic combined lift. Flips the current run rate of +$1.44/trade to +$3.30 to +$4.50 per trade average. The only meta item still open is the trade-#150+ full re-audit.

Phase-1 realized P&L (since Aug 21 ET)

Frozen snapshot. SSOT is scripts/phase1_pnl_record.json (prod trade_audit_log read-only SELECT 2026-09-26; filter = TRADE_METRICS_FILTER_SQL; net = PNL_EFFECTIVE_SQL; boundary = Config.PHASE1_START_ET / #367). Do not invent net-liq. Do not hand-edit the generated block.

SSOT: frozen scripts/phase1_pnl_record.json (prod read-only 2026-09-26). Net of commissions via PNL_EFFECTIVE_SQL (#331) except BBW id 274 (broker realized_pnl − recorded commission; stored net_pnl_after_financing is a stale $37.25 stop stamp). Boundary Config.PHASE1_START_ET (#367). Do not invent net-liq. Do not invent BBW's missing exit commission.

• Queried at (UTC): 2026-09-26T15:59:19.928748+00:00
• Phase-1 closed canonical trades: 13
• Phase-1 net (after commissions): $-922.28 — before missing BBW exit commission (NULL in ibkr_executions and trade_audit_log.exit_commission; not invented)
• Phase-1 gross (realized_pnl sum): $-912.23
• Phase-1 commissions: +$9.86
• Testing-era frozen (entry < Aug 21 ET): n=193 net +$194.16 (#367 post-deploy 193 closed; this snapshot n=193 net +$194.16)
• Aug 16–21 testing-era tail (not Phase 1): n=10 net +$73.35
• Open positions at query: 0
• trading_errors ledger: n=6 sum $-248.79 (no rows after Jul 31; #353)

IBKR-sourced capital (cited; not invented)

• NL Δ Aug 21 → Sep 26 snapshot: $-918.94
• Phase-1 canonical net: $-922.28 — before missing BBW exit commission (NULL in ibkr_executions and trade_audit_log.exit_commission; not invented)
• BBW stale-field understatement (corrected in this record, not in prod DB): +$160.58
• Residual (NL Δ minus corrected known net): +$3.34 plus missing BBW exit commission — UNVERIFIED

† BBW id 274 canonical net uses broker realized_pnl − recorded entry commission; stored net_pnl_after_financing remains −$121.52.

Flags

• VERIFIED: id 274 stored net_pnl_after_financing/gross_pnl = -$121.52 (hardware stop stamp $37.25). Broker fill (34.66-39.21)×62 = -$282.10. Canonical net in this record = -$282.4628 (realized_pnl − recorded commission 0.3628). Exit commission NULL — not invented. Prod DB was NOT updated on this branch; the twin repair ships as Finding #477 (guarded migration). (source: prod trade_audit_log id=274 + ibkr_executions 550/555 + trading_positions id=307 read-only 2026-09-26; FINDINGS #423; correction: Finding #477)
• INFO: id 268 hardware stop is already canonical stop_loss:hardware in stats. Not moved to trading_errors (would need operator approval / #353). (source: prod trade_audit_log id=268; FINDINGS #353)
• INFO: Aug 16–21 testing-era tail n=10 net +$73.35 is NOT Phase 1 (entry < PHASE1_START_ET). (source: Config.PHASE1_START_ET default 2026-08-21 ET (#367))
• INFO: trading_errors n=6 sum -248.79; no rows after Jul 31 (#353). (source: prod trading_errors SELECT 2026-09-26; FINDINGS #353)
• UNVERIFIED: IBKR NL Δ Aug 21 $27,504.38 → Sep 26 snapshot $26,585.44 = -918.94 vs corrected Phase-1 net -922.2826; residual +3.3426 plus missing BBW exit commission. Without Flex the residual is not a cash proof. (source: FINDINGS #366; prod ibkr_account_snapshot 2026-09-26T16:11:47Z)
• INFO: Other-row scan: /realized_pnl−gross_pnl/>1 n=1 (BBW 274 only). /realized_pnl−net_pnl_after_financing/>1 n=2 (BBW 274 this class; SNOW 75 testing-era different class, not Phase 1). (source: read-only SELECT via prod droplet venv 2026-09-26)

2. Ranked action queue — single combined source of truth

Playbook §2 current queue and the dashboard 🎯 Action Queue are both produced by scripts/strategy_performance_analysis.py:build_action_queue() (memory.md §K/N). Edit that function, never a surface. The generated block is marked BEGIN GENERATED: action_queue. Live evidence numbers in the dashboard are recomputed on every report run from the same DB the trader uses.

🔬 Aug 16–17, 2026 — #357 full re-validation executed + #360 decision-state reconciliation + #361 approvals (READ THIS FIRST; supersedes the queue state below where marked)

The trade-#150+ audit trigger FIRED and was executed as Finding #357 (n=183; docs/STRATEGY_REVALIDATION_2026_08.md, re-issued as Parts I–IV). The operator then decided (Part I — Finding #360, Aug 16) and APPROVED the three open Part II decisions on Aug 17 (Finding #361) — every #357-program decision is now made. The authoritative current queue is the dashboard Action Queue + the

357 section's ordered action plan; summary:

• DECIDED Aug 16 — IMPLEMENTED Aug 20 (#366), DEPLOYED Aug 21, 2026 (rank-1 package package-10k-s2c-b4-357; SoT: FINDINGS #366): C-Q1 trading capital → $10,000 under the HARD PDT constraint (account equity ≥$25k at ALL times; reserve ≈$17,450 NOT withdrawn; no day-trade counting exists in code) + S2c sizing (enter everything / equal split / shrink-to-fit / ≤25% per position) + the B-4 3-tier priority table (rank 3 = LONG 70–75 any mcap · 2 = LONG 80–85 large/mid · 1 = rest · 0 = skip; all volume bumps + the mcap spine removed — F-001/F-002 refuted). The package's decision set is COMPLETE (C-1 + RES approved Aug 17 — #361); ships as ONE reviewed package on the operator's EXPLICIT go; prerequisite fix A-7; arms the K-S2C/K-B4/K-RES/M-C1 (+K-B5) re-check contracts (all alert-only — self-alerting since #389/#392, the operator decides every outcome). Changelog entries at SHIP time (#338 convention).
• Implemented + DEPLOYED (Aug 16, 2026 20:22 UTC): #358 (PM print-density collector fixes = A-8)

• 359 (B-3 stale-signal guard; closed A-4) + #360 (this reconciliation) — single

  deploy_from_github.sh run at efebe45; reaper closed 1,288 stale rows on first run.
• APPROVED Aug 17, 2026 (#361) — IMPLEMENTED Aug 20 (#366), DEPLOYED Aug 21, 2026: C-1 re-close of the 0.90–0.95 LONG band (recommendation CLOSE accepted; §14 settled the contest; M-C1 shadow monitor ARMED) · RES the M1+M5 reservation mechanism (observed-set + rank-3 AMC carve-out, multi-day component dropped; K-RES ARMED) · B-5 small-cap LONG 0.80–0.95 pre-registered pilot (K-B5 ARMED, §12.5 revert terms unchanged — alert-only since the #392 amendment, Aug 21: the operator decides the revert). All three are live in production.
• Still OPEN (strategic, time-gated — not part of the package): C-Q3 (Oct capacity). C-Q2 DECIDED Sep 26, 2026 — keep (#478; next at the count trigger).
• Queue effects below: the "Re-validate full audit at trade #150+" row is DONE (#357); §7's VIX true-sim ask is DISCHARGED (D-2 deep dive — no rule ships; shadow monitor instead); the #314.e SKIP candidates are RETIRED (superseded by B-4); v3 D is folded into B-4; v3 E re-keys to the D-5 SHORT re-test (≥40 CF cases) / B-008 (≥20). All 17 armed monitors/triggers/contracts (16 from #360/#361 + D-9 #363; SoT: build_revalidation_357()['monitors']) live in ONE table: STRATEGY_REVALIDATION_2026_08.md §II.7 (KILL, D-1…D-9, M-C1, M-VOL, K-S2C, K-B4, K-B5, K-RES, G-RS), rendered on the dashboard.

SSOT: scripts/strategy_performance_analysis.py:build_action_queue(). Do not hand-edit this block. History tables below stay as dated narrative.

Generated from 50 Action Queue rows.

IN-QUEUE (1)

ON-HOLD (8)

IMPLEMENTED (21)

FAILED-GAUNTLET (3)

FAILED-REGAUNTLET (1)

RETRACTED (10)

KEPT (3)

INFO (3)

Historical queue tables (superseded — generated Action Queue above is SSOT; history)

🟢 IN-QUEUE — what to ship next (in order)

Apr 25 LATE evening updates (Finding #232 + #233 re-validation): - v3 Finding F (BMO + VIX sizing) DROPPED to RETRACTED — deep bias gauntlet showed 1.2× scale-UP arm fails ex-top-3 (+$100 concentrated in STT/VZ/ALLY); direct modeling gives only +$36/100 not +$60; operator dropped both arms. - RuleB PROMOTED rank 4 → 2 — corrected stop classification (Finding #233; old stop_triggered field undercounted at 1.7 % vs real 49 %) revealed stops are bleeding −$676 / 57 stopped trades, the strategy's biggest leak. RuleB is the only stop-loss rule that survives the gauntlet. (SHIPPED May 28 as Phase-1 DRY-RUN — Finding #305.) - v2 #6 (3-stop session breaker) confirmed has historical fire — 9 sessions of ≥3 stops in history (Feb 19 had 5); was wrongly believed to have zero historical evidence. (SHIPPED May 30 — Finding #312.) - v3 Finding D (Mid-cap 80%+ skip) REASONING REINFORCED — Mid 80-90% has 52 % stop rate, Mid 90-95% has 86.7 % stop rate (massacre). Stops-eat-winners hypothesis vindicated; KEEP RuleB dependency. - v3 Finding E (non-downtrend SHORT skip) REASONING REINFORCED — sideways SHORT 83.3 % stop rate (was claimed 67 %). RuleB likely fixes; KEEP dependency. - 5-outlier curse (renamed Apr 25 from 3-outlier — Finding #232 + memory §K): GPRO + DELL + SWBI + TXN + STUB carry ~$430 of $167 net.

Apr 25 earlier evening updates (kept for context): - v3 Finding C (drift filter) RETRACTED via option 3 (drop and rely on B+D+E + selection rules; universe view weakened the bias check).

✅ SHIPPED Jul 11, 2026 — busy-week capital bundle (Finding #335)

Operator-directed after the Jul-14-week deep-dive (bank wave: 3+/day expected, overlapping AMC/BMO holds — busiest week on record). (1) Exposure cap 1.40 → 1.50 (MAX_DAILY_EXPOSURE + unwired MAX_TOTAL_EXPOSURE_PCT lockstep): flat-25% sizing fit only 5 positions (125%) under 1.40 — 15%/night structurally stranded; 1.50 fits exactly 6. Targeted granularity fix; NOT the broad 2.0 raise #315 declined. Margin-safe (6×$676 ≈ $4.05k vs enforced 2.0× Reg-T gate ≈ $5.4k). (2) Admission freed-capital fix: can_open_position now nets out positions exiting before the new entry time — pre-fix a full overnight book auto-rejected EVERY next-morning AMC ("would exceed cap" at 10:30 although the book exits 10:32); Jan-28 fix covered batch sizing only. (3) BMO reservation re-based: available capital = strategy budget bounded by IBKR BP (was RAW day-trade BP ~$110k → reservation never bound). Session breaker (#312) threshold RAISED 3 → 4 (#335.b, operator choice): 9-day counterfactual — only 4 would-be-blocked entries, net −$15.78 (breaker would have SAVED money); agent recommended keep-as-is, operator chose 4 to tolerate one more correlated stop on 6-slot days (skip-reason string session_3_stop_breaker kept for analytics continuity). Monitor Jul-14 week: 6-slot nights, first-ever exposure_cap skips, AMC admissions on full-book mornings, breaker interplay on the correlated Tuesday bank stops.

✅ SHIPPED Jul 11, 2026 — #202 LONG cap RELAXED ≥0.90 → ≥0.95 (Finding #333) + priority re-tiered 0-10 (Finding #334)

Operator chose the outright relax (over half-size pilot / keep). SIGNAL_LONG_MAX_THRESHOLD 0.90 → 0.95 (config default + prod .env); lockstep per §M: conviction_calculator.is_high_conviction, simulator DirectionPolicySpec (prob_threshold_70_30_with_95_skip), capital_manager._bmo_reservation_threshold (0.95 — preserves pre-#333 effective no-override behavior), dashboard Python + Jinja + JS gating (badges now ✗95+ / 70-94%), Live Config Direction row, P&L-progression Jul-11 marker, STRATEGY.md ×8, Action Queue rows, 2 guard tests. ≥0.95 skip REMAINS.

⚠️ Finding #334 (same day, operator-caught): the #333 rollout MISSED priority.py, which independently zeroed ≥0.90 LONGs — the reopened band would have been silently dropped at priority_filter. Fixed with the operator-specified re-tier: base 0-8 (Large 70-80/80-90/90-95 = 8/7/6 · Mid = 5/4/3 · Small 70-80 = 2, 90-95 lowest per mcap group, counterfactual-only evidence) + #313 volume bump → 0-10 scale (MAX_PRIORITY_SCORE=10). Small caps stay 70-80% LONGs only; SHORT miss-conf ≥90% stays SKIP. Dashboard legend/badges/tooltips all on /10. Lesson (§M, then-live): the LONG cap was encoded in BOTH Config. SIGNAL_LONG_MAX_THRESHOLD AND priority.py:LONG_MAX_CONVICTION — change together. (#399, Aug 22: priority.py now reads Config.SIGNAL_LONG_MAX_THRESHOLD lazily — the numeric twin is gone; changing the env/config is enough.) Monitor: first ~20 real fills in the reopened band vs the +$6.19/trade counterfactual; re-gauntlet on REAL fills at trade #150 or 20 band-fills. The pre-ship evaluation record follows.

👤 evaluation record — relax #202 to 0.90-0.95? (Jul 10, 2026 — Finding #333)

Operator asked (Jul 10) whether the ≥0.90-skipped markets would have been winners — answered with a gap-aware 1m-bar replay (scripts/ analyze_90_95_bucket.py) of all 112 #202-skipped markets instead of a live half-size pilot: scheduled entry/exit bars, per-tier 3/4/5% stops active 24/7 with gap-aware fills (#330 lesson), flat-25% sizing (#314.f), net of commissions.

Band	n	Net (Tiered)	ex-top-2	WR	stops	Verdict
0.90-0.95	78	+$482.94 (+$6.19/tr)	+$135.39 PASSES	42%	51%	operator decision
0.95+	34	−$6.03	−$228.48	32%	62%	KEEP skipping

Fragility (per §K, reported honestly): sub-slices mostly FAIL ex-top-2 (large −$157, mid −$55, pre-May −$95; only small-caps + the pooled band pass); ~2/3 of recorded convictions pre-date the #325/#327 estimator fixes; no slippage modeled; wrong-prediction instant-exit rescue NOT modeled (conservative). Contrary evidence #323.b (26 real pre-Mar-9 fills −$112) used the buggy estimator + old rule set. Options: (a) relax cap to 0.95, (b) relax at half-size until ~20 real fills, (c) keep #202, re-run at trade #150+. Mirror row: relax-202-9095. Same finding also FIXED the exclusion-funnel coverage leaks (channel-window hole + gamma-live conviction fallback + end_time timing inference).

✅ SHIPPED Jul 6, 2026 — sizing-up re-validation → FLAT 25/25/25 (Finding #314.f)

Jul 6, 2026: operator chose flat 25/25/25 (over 35/25/18 and KEEP) via AskQuestion — SHIPPED to production (config.py POSITION_SIZE_PCT_LARGE/MID/SMALL 0.30/0.22/0.15 → 0.25/0.25/0.25; simulator rules.py sizing spec jul6_flat25 + production_strategy_jul6() factory in lockstep; dashboard Live Config + STRATEGY.md + memory.md updated). Monitor small-cap stop-outs (now sized 25% in the 5%-stop tier) over the next ~20 trades; re-validate at trade #150+. The pre-ship evaluation record follows.

Jul 5, 2026 capacity-aware re-run of the parked #314.c sizing-up candidates (replay_sizing_variants.py --exposure-cap 1.40, 578 signals = 134 live + 444 counterfactual, 39.0 weeks, June trades included). Then-baseline (since superseded) 30/22/15 = +$20.99/wk. Both raise/concentrate candidates gauntlet-PASS and STRENGTHENED vs May 31:

Variant	Δ/wk	Total	WR	maxDD	5p	Gauntlet
flat 25/25/25	+$7.40 (was +$4.18)	+$1,107.04	44.0 %	−$245.46	−$98.60 (unchanged)	✅ PASS
35/25/18	+$6.08 (was +$4.63)	+$1,055.71	44.4 %	−$236.88	−$113.65 (worse)	✅ PASS
33/22/15 (Large only)	+$0.10	—	—	—	—	❌ FAIL-EX-TOP-2
all lower-sizing variants	negative	—	—	—	—	❌ FAIL-SPILLOVER

Caveats unchanged from #314.c: flat 25/25/25 raises Small 15 %→25 % (squeeze risk in the 5 %-stop tier); 35/25/18 raises Small only to 18 % but has the worse 5p tail. Operator decision pending: ship flat 25/25/25 / ship 35/25/18 / KEEP the then-live 30/22/15 (since superseded by flat-25 #314.f, then S2c #366; re-check at trade #150+). Mirror row: build_action_queue() id sizing-up-2026-07.

👤 PENDING-OPERATOR-APPROVAL — RESOLVED May 30, 2026

Finding #311 SHIPPED gate_uniform_$1000 to production (May 30, 2026 LATE evening). First real-money strategy ship since Mar 9. After full correctness audit caught + fixed 6 errors in earlier simulator outputs, the corrected per-tier empirical scaling showed gate_uniform_$1000 is Pareto-optimal: Δ E[wk] +$21/wk (~+$1,092/yr at $2,450 capital), Δ max DD -$46 (within -$735 target = 30% capital), outlier alarm fixed. Production code: src/earnings/config.py SIGNAL_MIN_VOLUME_TOTAL default 2000.0 → 1000.0. production_strategy_may30() factory captures new state. The other 2 PR 2 ship-eligible variants (gate_quarterly_$2000peak_$1000other, gate_quarterly_$1000peak_$500other) RETRACTED-IN-FAVOR-OF gate_uniform_$1000 — operator picked simplest single-axis change. gate_uniform_$250 (more aggressive) DEFERRED to monitor uniform_$1000 effects first.

⏸️ ON-HOLD — paused pending another rule or more data

Jul 5, 2026 (Finding #330): the RuleB dependency on both rows below is REMOVED — RuleB failed the gap-aware gauntlet and its Phase-2 active flip is CANCELLED, so the stop bleed these rules were waiting on has NO shippable stop-side remedy. Both re-evaluate as PURE selection rules.

✅ IMPLEMENTED — historical rules already live (kept for traceability)

🛑 RETRACTED / ⚠️ FAILED-GAUNTLET — kept inline so future audits don't re-suggest them

ℹ️ INFO — informational, no action needed

3. Verified findings — mechanism, evidence, bias check

For each rule, the documentation pattern is: mechanism (why it should work) → evidence (the table) → bias check (4-check gauntlet result).

#1 — Skip LONGs in downtrend regime (v3 Finding B / v2 #1) — ✅ IMPLEMENTED May 28, 2026 (Finding #304)

Ship status (May 28, 2026): SHIPPED as first Phase-3 strategy rule after the May-26 un-gate. Commit lives at src/earnings/signal_generator.py (new private _safe_get_regime() helper + skip block inserted in _build_trading_signal() after the small_cap_long_above_80 check). 7 new unit tests in tests/unit/test_signal_regime_gate.py (LONG×downtrend skip, plus 6 fail-soft / direction-asymmetry / regime-variation negative-controls). Live acceptance test: monitor polymarket_opportunity_log.skip_reason='long_downtrend' rows over the next downtrend regime. Note: at current PM volume-gate pass rate of 1.9 % (Finding #279 amendment), natural firing rate is ~0.4 % of all earnings markets — live "≥30 trades" gauntlet trigger may take years; rule's immediate value is in bias-gauntlet-locking the policy + correct behavior in the eventual next high-volume regime, NOT in fast live-evidence accumulation.

Mechanism: Earnings beats matter less when the broader market is selling. Even high-conviction LONGs get caught in macro flows; the stop fires on the macro move, not the earnings reaction.

Evidence (refreshed Apr 25 evening, 116 closed trades):

Bias check (4 checks) — STRONGER after Apr 25 re-validation: - ✅ n=22 (≥12 minimum); 22/22 distinct tickers (no repeats — bias #3 trivially passes) - ✅ Ex-top-2 by |pnl| = −$175 (more negative than headline! The top-2 by |pnl| are SWBI +$70 winner + a smaller loser; SWBI was masking how bad the bucket really is) - ✅ Top losers across 8+ distinct tickers (BABA, GME, KSS, NKE, PDD, ASAN, VSCO, BBY) - ✅ Full-population spillover positive (uptrend +$71, sideways flat)

Stop rates corrected via exit_reason ILIKE '%stop%'; the old stop_triggered boolean undercounted at 1.7 % across all trades and is broken (Finding #233).

Implementation: Skip rule on signal_generator using MarketRegimeProvider.get_snapshot().regime. ~10 lines.

#2 — Skip LONGs with pre-event drift ≥ 2 %

Mechanism: Stocks that have already drifted into our LONG direction overnight have priced in the EPS expectation. The small additional move on the announcement isn't enough to cover commission + slippage + stop risk.

Evidence:

Bias check: - ✅ n=15 in the skip bucket (combined 2-5 % + ≥5 %) - ✅ 14/15 are losers; ex-best-1 (EBAY +$5.64) still −$150 across 14 - ✅ Monotonic — more drift, worse outcome (no cliff edge) - ✅ Full-population: skipping 15 of 90 LONGs preserves the +$107 winners

Implementation: Use earnings_markets.stock_price_day_before_close (populated by Finding #227 resolver on 713/715 past markets); compute drift = (current − stock_price_day_before_close) / stock_price_day_before_close at signal time. ~15 lines.

#3 — Skip Mid-cap LONGs at conviction ≥ 80 % — ⏸️ ON-HOLD (RuleB dependency REMOVED Jul 5, 2026 per Finding #330 — re-evaluate as pure selection rule)

Mechanism: High-conviction Mid-caps run the most into earnings (smart money front-running) so the move is most pre-priced. Large caps have deeper books and less front-running pressure. Apr 25 alternate mechanism (vindicated): stops eat winners on Mid-cap (Mid 80-90 % has 51.6 % stop rate, Mid 90-95 % has 86.7 % stop rate) — RuleB likely fixes.

Evidence (refreshed Apr 25 evening with corrected stop classification):

Bias check: - ✅ Both Mid sub-buckets independently negative across n ≥ 15 - ✅ Mid is 70 % of trade count — high-leverage rule - ⚠️ Large 80-90 % degraded from +$22 to −$3 — the "Mid-only" framing is weakening; potential future widening to "Skip ALL LONGs ≥ 80 %" - ⚠️ Universe view at $2K gate shows Mid 80-90 % is POSITIVE without stops (Finding #228 stops-eat-winners hypothesis) — RuleB resolution highly likely

#4 — Skip SHORTs in non-downtrend regimes — ⏸️ ON-HOLD (RuleB dependency REMOVED Jul 5, 2026 per Finding #330 — re-evaluate as pure selection rule at SHORT n≥30)

Mechanism: SHORTs are a different bet entirely — they need market weakness to amplify the bad-earnings move. In uptrend / sideways the recovery rally bounces our short out. Apr 25 alternate mechanism (vindicated): sideways SHORT 83.3 % stop rate (corrected from 67 %) strongly supports the stops-eat-winners hypothesis — RuleB likely fixes.

Evidence (refreshed Apr 25 evening with corrected stop classification):

Bias check: - ✅ 13/13 negative on the skip side — perfect direction signal - ⚠️ Downtrend-positive KEEP set is 2-of-3 outliers per the 5-outlier rule (GPRO + STUB; ex-them just GIS +$3.64) - ⚠️ Total SHORT n=18 below the n≥30 commit threshold

#5 — Daily 3-stop session breaker (v2 #6 carryover) — ✅ IMPLEMENTED May 30, 2026 (Finding #312)

Mechanism: Tail-risk cap. After 3 stops in a single trading session, something systemically bad is happening (regime shift, earnings-day flash). Bot pauses entries until next session.

Evidence: 9 historical sessions with ≥3 stops per the May-30 #312 production query (2026-02-03 → 2026-04-29). The Apr-25 re-validation found 8 — including Feb 19 (5 stops, −$35), Mar 2 (4 stops, −$59), and six 3-stop days; total stops on those 8 days: 27, costing −$266. Was previously believed to be tail-hedge only with no historical fire — that claim was an artifact of the broken stop_triggered field (Finding #233).

Implementation (AS SHIPPED, #312): pure-decision module src/earnings/session_stop_breaker.py + _count_todays_stops() helper + skip block in signal_generator._build_trading_signal (after Queue Rule #1 long_downtrend); counts session stops via trade_audit_log (using exit_reason ILIKE '%stop%', NOT stop_triggered); if ≥threshold today (SESSION_STOP_BREAKER_ENABLED=True) → log skip_reason='session_3_stop_breaker' and skip. Threshold raised 3 → 4 Jul 11, 2026 (Finding #335.b, operator choice for the busy week) — the skip-reason string is unchanged for analytics continuity.

#6 — BMO + VIX-bucket sizing (v3 Finding F) — 🛑 RETRACTED Apr 25

Moved to §4. Apr 25 LATE evening (Finding #232) bias gauntlet split verdict: scale-DOWN arm passes (BMO+VIX 18-22, n=30, −$54 robust); scale-UP arm fails ex-top-3 (BMO+VIX <18 +$100 collapses to −$1.58 across 17 ex STT/VZ/ALLY). Direct modeling shows total per-100 lift is +$36 (NOT +$60), and that's optimistic. Operator dropped both arms. Re-test only the scale-DOWN arm at trade #150+ with fresh data.

#7 — RuleB (±90 min no-stop window around earnings release) — 🛑 FAILED-GAUNTLET Jul 5, 2026 (Finding #330); Phase 1 DRY-RUN was REMOVED Aug 21, 2026 (#380; was logging-only)

🛑 Jul 5, 2026 REVERSAL (Finding #330): the evidence below was an artifact of the simulator's gap-fill optimism — simulate_window_only filled window-blocked stops at the STOP PRICE even when the price was still beyond the stop at rearm. Gap-aware (rearm fills at the first post-window bar OPEN, matching the as-shipped cancel-and-rearm order semantics), RuleB is −$209/100 headline / −$194/100 ex-top-2 / −$227/100 ex-repeat (n=129) — FAILED-GAUNTLET. Phase-2 active flip CANCELLED. The section below is preserved as the historical record of the pre-#330 reasoning.

Ship status (May 28, 2026; history): Phase 1 SHIPPED in DRY-RUN mode. Code lived (was deleted #380) in src/earnings/ruleb_window_gate.py (pure decision function + RuleBAction enum + RuleBDecisionContext dataclass) plus dispatcher methods _process_ruleb_window_gate / _ruleb_fire_cancel / _ruleb_fire_rearm on trading_orchestrator. Hardware cancel-and-rearm pattern (NOT soft-gate per Finding #228 lesson — the simulator's simulate_window_only requires hardware-level silencing; soft-gate would deliver $0 EV like FA_soft did). Phase 1 RULEB_MODE='dry_run' default = logs cancel/rearm actions WITHOUT IBKR side effects. Phase 2 (separate later session) flips RULEB_MODE='active' after operator reviews dry-run logs. Phase 1 known limitation: restored positions (signal=None per position_manager.load_persisted_positions:1155) cannot derive earnings_release_dt → graceful SKIP_NO_EARNINGS_DT log; Phase 2 may add restart-safe lookup if dry-run shows this happens in practice. 25 unit tests in tests/unit/test_ruleb_window_gate.py cover all 8 RuleBAction return values + sanity-gate skips + late-entry edge case + hard-fallback safety boundary; 36/36 PASS pre-commit (25 RuleB + 11 prior signal-test regression).

Mechanism: Earnings release momentarily widens spreads and triggers stops on volatility, not on direction. The ±90 min window covers the legitimate volatility spike.

Evidence (Apr 25 simulator): +$62.97 vs A_base across 51 stops (+$123 / 100 trades). 4 better / 1 worse / 46 unchanged → 96 % no-effect. Apr 25 evening re-validation confirmed live stop frequency matches simulator population: 57 of 116 trades = 49 % stop rate, costing −$676 (corrected via exit_reason; old stop_triggered field reported only 1.7 % — Finding #233). Stops are the strategy's biggest leak; non-stops contribute +$843.

Bias check: - ✅ n=51 - ✅ Robust to ex-top-2 (still positive at smaller magnitude) - ✅ Survives ex-repeat-tickers (no name dominance) - ✅ Full-population spillover: 96 % unchanged trades, no spillover risk

Implementation: Software gate in order_executor.py on stop-fill event handler: if (now - earnings_release_dt).abs() < 90 min, log "stop blocked: in earnings window" and don't act on the fill (let the hardware order fill, then immediately re-enter at scheduled exit if position is still flat). ~25 lines.

PROMOTED rank 4 → 2 on Apr 25 evening because the corrected stop classification revealed live stop bleed dominates the negative contribution. Rule #1 SHIPPED May 28, 2026 (#304); RuleB shipped right after it the same day as Phase-1 DRY-RUN (#305).

4. Retracted / failed findings — for the historical record

Documented here so future audit passes don't re-discover them.

🛑 v3 Finding A — Block stops outside RTH (RETRACTED Apr 25, Finding #228)

Original claim: +$300–$400 / 100 trades, LOW risk, NO selection change. Implement first.

Why retracted: Comprehensive simulator (scripts/analyze_finding_a_full.py):

• FA_soft (rule as written): $0.00 vs A_base. The hardware STP-LMT carries outsideRth=True per Finding #179, so a software-only gate is literally inert — the broker still fills regardless of bot bookkeeping.
• FA_hard (hardware also gated): −$213.23 vs A_base across 51 stops (−$418/100). The "rearm_gap_open" mechanism — stop blocked in extended hours, RTH opens past the stop, exit forced at gap-open price — costs $417 across 22 trades, outweighing the 9 recovery wins (+$204) by 2-to-1.

Audit error: Used Rule B (±90 min window) as a proxy and treated the per-session observational table as counterfactual evidence. Rule B is narrower than Finding A and the per-session table just measures where stops fired (tautologically negative because stops cause losing exits).

Mechanism failure: The audit's "stocks fill at bad prices then revert during RTH" assumption isn't supported. For 22 of 51 stopped trades the stock kept falling past the stop into RTH open. Pre-EPS-trigger only (the audit's strongest claimed scenario) is also net-negative at −$92.

🛑 v3 Finding F — BMO + VIX-bucket sizing (RETRACTED Apr 25 LATE evening, Finding #232)

Original claim (v3 audit Apr 22): +$60 / 100 trades by scaling 1.2× / 0.6× / 1.0× across BMO + VIX <18 / 18-22 / ≥22 buckets. Robust on both ends per audit's ex-top-3 check.

Why retracted (deep bias gauntlet on Apr 25 evening with fresh data):

Direct multiplier-on-realized-pnl modeling: Baseline BMO total +$39.72 → After F (full 1.2/0.6/1.0): +$81.24 → delta +$36 / 100 trades, not the docs' +$60. And that's optimistic (assumes linear scaling, ignores commission/slippage fixed costs and capital constraints at ~$2.4k).

Operator decision Apr 25 LATE evening: Drop both arms. Scale-DOWN arm alone is only ~$20/100 lift — not worth the implementation + variance + capital-constraint risk at current capital. The scale-UP arm is fragile and adds risk for marginal gain.

Re-test trigger: Trade #150+ with fresh data. Only re-introduce the scale-DOWN arm if ex-top-2 stays robustly negative. Skip the scale-UP arm entirely until the BMO+VIX <18 winner bucket survives ex-top-3.

🛑 v3 Findings D-stop / G-stop — ATR-based stops (FAIL gauntlet Apr 25)

Both collapse to net-negative when the top-2 outliers (KSS / GME / BBY recovery cluster) are removed. Bias-check #2 fails. Could be revisited when the 3-outlier dominance dilutes (n ≥ 100 stops).

🛑 v2 #4 — VIX 22+ is the best bucket (RETRACTED v3 audit)

64 % of bucket P&L from GPRO + SWBI. Ex-them: +$36 across 12 trades. Replaced by Finding F (BMO + VIX <18) which is bias-robust.

🛑 v2 #6 — L 70-80 % is the only profitable LONG band (RETRACTED v3 audit)

82 % of bucket P&L from SWBI. Ex-SWBI: +$15. Ex-SWBI+DAL: −$17. Kept only as "don't penalize" baseline (i.e. don't add a skip rule on this band — it isn't actually a sell signal).

🛑 v2 H-rule — ATR > 5 % pre-screen (RETRACTED v3 audit)

The "smoking gun" for full-population spillover: headline +$329 on the 49-trade stopped+correct subset → −$148 net across 109 trades when applied to the full population. Skip rules in particular need the full- population check because they remove WINNERS as well as losers.

5. Mandatory bias gauntlet (memory §K)

Every new finding / rule MUST pass all four checks below before being added to the queue. Source: docs/STRATEGY_AUDIT_v3.md §1 + §6 + .cursor/memory.md §K.

1. n adequate — n ≥ 12 minimum, ideally n ≥ 20.
2. Ex-top-2 outliers — drop the two largest |delta| trades. Finding must still survive in direction and magnitude.
3. Ex-repeat-tickers — de-dupe by unique ticker (most-recent kept). Finding must survive when repeated evaluations of the same name count once.
4. Full-population spillover — apply the rule to ALL trades, not just the targeted subset. Skip rules in particular need this — they remove WINNERS as well as losers.

Plus the 5-outlier rule (renamed Apr 25 evening from 3-outlier — Finding #232): GPRO + DELL + SWBI + TXN (added Apr 22, +$92, LONG large NULL-regime) + STUB (added Mar 4, +$58, SHORT mid downtrend) together carry ~$430 of the strategy's ~+$212 net P&L on ~117 closed trades; ex-top-5 the strategy is −$263 across 111 trades. ANY slice that includes them looks great; ANY slice that excludes them is structurally negative. ALWAYS report numbers ex-top-2 alongside the headline; for any slice containing 2+ of the named 5 outliers, also report ex-them.

Plus the date-drift rule (added Aug 18, 2026 from Finding #362): any analysis that joins historical trades to earnings_markets.earnings_date (entry-day correctness, wrong-date-entry counts, release-window studies) must treat pre-#349-era rows (before Jul 29, 2026) as unverified at trade time — sync writers moved earnings_date AFTER events, so today's stored date is not the date the trade decision saw. Known false positives: ON (Feb-9) and NET (Feb-10) LOOK wrong-dated in today's DB but entered on the correct day (the DB date drifted later post-event). Before classifying any historical entry as wrong-dated, cross-check trade_audit_log entry dates against earnings_date_audit history. The historical rows are deliberately NOT repaired (repairing would falsify the as-traded record). The genuine wrong-date class (4 of 193: CRSP, SBUX ×2, VIK) is closed forward by the

362 guards.

Plus the true-simulator rule (added Apr 25 from Finding #228): every counterfactual rule needs a TRUE simulator that replays the proposed rule end-to-end on the actual data. Proxies (e.g. "Rule B is close enough to Finding A") and observational tables (e.g. "stops fire pre-market AND pre-market trades have negative P&L → blocking pre-market stops would have helped") are NOT counterfactual evidence — they conflate selection with causation.

6. Implementation notes per rule (file + LOC + acceptance test)

When implementing each rule, follow memory §B (5-step verification) and §H (commit + push + deploy). Each rule below has a pre-baked 5-step trace to copy into the implementation PR.

#1 — Skip LONGs in downtrend regime — ✅ IMPLEMENTED May 28, 2026 (Finding #304)

Implementation notes preserved for historical context + future reference. Actual code matches the spec below with minor refinements (insertion after small_cap_long_above_80 instead of "before volume gate" per operator AskQuestion choice; added _safe_get_regime private helper rather than inlining). See docs/FINDINGS.md Finding #304 for the canonical narrative.

File: src/earnings/signal_generator.py (helper _safe_get_regime ~17 lines + skip block ~22 lines) + tests/unit/test_signal_regime_gate.py (7 unit tests).

Step 1 — Trace path (AS SHIPPED): _build_trading_signal() → conviction direction classification → volume gate (may reject) → get_market_cap_for_symbol + classify_market_cap → small_cap_short skip (no-op for LONG) → small_cap_long_above_80 skip (no-op unless Small+≥80 %) → [Rule #1 here] if direction == LONG: current_regime = self._safe_get_regime(); if current_regime == "downtrend": return (None, "long_downtrend") → callers (BMO/AMC loops, line 1606 / 2052) see reject=="long_downtrend" → _emit_pm_opportunity_log({skip_stage: "generate_signal", skip_reason: "long_downtrend", ...}) → no entry order placed.

Step 2 — Simulate (AS SHIPPED): at signal time on a Mon BMO trade for AAPL when SPY is in downtrend (per MarketRegimeProvider._classify reading SPY MA20/MA50/MA200): bot calls _build_trading_signal() → conviction 0.82 (LONG) → volume_total $5,000 (passes $1k gate, #311) → mcap=Large → small_cap_short skipped (LONG, not SHORT) → small_cap_long_above_80 skipped (Large, not Small) → Rule #1 fires: _safe_get_regime() returns "downtrend" → logger.info("...Queue Rule #1...") → return (None, "long_downtrend") → caller emits opportunity_log row with skip_reason="long_downtrend" → continue (no order).

Step 3 — Line numbers (AS SHIPPED, source as of May 28, 2026): - Helper: src/earnings/signal_generator.py:_safe_get_regime (immediately after _regime_fields_for_log at line ~1067). - Skip block: src/earnings/signal_generator.py:_build_trading_signal (after small_cap_long_above_80 check at line ~1213, before calculate_entry_exit_times at line ~1234). - Tests: tests/unit/test_signal_regime_gate.py (new file, 7 tests).

Step 4 — Failure modes (AS HANDLED): - _regime_provider is None (test mocking bypasses init, OR production instance built before regime was wired in) → _safe_get_regime() returns None → no skip (preserved current behavior). Covered by test test_long_with_no_regime_provider_is_not_skipped. - provider.get_snapshot() raises → caught at DEBUG (mirrors _regime_fields_for_log pattern) → returns None → no skip. Covered by test test_long_when_regime_provider_raises_is_not_skipped. - snap.regime == "unknown" (classifier can't determine state, e.g., insufficient SPY MA history) → does not equal "downtrend" → no skip. Covered by test test_long_with_unknown_regime_is_not_skipped. - snap.regime == None → does not equal "downtrend" → no skip (same code path as "unknown"). - Stale regime cache (≤ 60s per MarketRegimeProvider _CACHE_TTL_SECONDS) → at most 60s drift between actual SPY state and classifier output → acceptable; the rule is statistical not deterministic. - SHORT direction with downtrend regime → rule does NOT fire (LONG-only) → preserves the historical +$172/3-trade downtrend-SHORT bucket. Covered by test test_short_in_downtrend_is_not_skipped_by_rule1.

Step 5 — Acceptance test (AS RUN): - ✅ pytest tests/unit/test_signal_regime_gate.py -v — 7/7 PASS (verified pre-commit on prod's Python 3.12 venv). - ✅ Broader signal-test regression: pytest tests/unit/test_signal_volume_threshold.py tests/unit/test_signal_reliability_gate.py tests/unit/test_signal_regime_gate.py — 11/11 PASS. - Live post-deploy: journalctl -u earnings-trader --since "1 hour ago" | grep -E "(Queue Rule #1|long_downtrend)" should match for any LONG signal generated when SPY is in downtrend. May be zero matches at current volume regime (1.9 % gate-pass) — that's expected; the rule's value is policy-locking, not fast live evidence. - Live post-deploy SQL: SELECT skip_reason, COUNT(*) FROM polymarket_opportunity_log WHERE evaluated_at > NOW() - INTERVAL '7 days' GROUP BY skip_reason ORDER BY 2 DESC — long_downtrend should appear if any LONG signal generated when SPY in downtrend during the window.

#2 — Skip LONGs with drift ≥ 2 %

File: src/earnings/signal_generator.py (~15 lines).

Step 1 — Trace path: _build_trading_signal() → after volume gate → fetch earnings_markets.stock_price_day_before_close for this condition_id → fetch current quote (use _pre_signal_quote() helper) → compute drift_pct = (current − prior_close) / prior_close → if direction == "LONG" and drift_pct ≥ 0.02, skip with skip_reason="pre_event_drift_long".

Step 2 — Simulate: at 15:58 ET prior-trading-day for a BMO trade on WDAY when prior close was $250 and current is $258: drift_pct = +0.032, direction LONG → skip.

Step 3 — Line numbers: insertion point = after volume gate, before final signal construction.

Step 4 — Failure modes: - stock_price_day_before_close is NULL on this market → CONTINUE (don't skip; the resolver covers 713/715 past markets but new markets may be NULL until the daily 06:00 ET resolver runs). - Quote fetch fails → CONTINUE. - Drift exceeds threshold but direction is SHORT → unchanged (rule only applies to LONG).

Step 5 — Acceptance test: - Mock earnings_markets.stock_price_day_before_close = 100, current quote = 103, direction LONG → expect skip. - Same with direction SHORT → expect signal generated.

#3 — Skip Mid-cap LONGs at conviction ≥ 80 %

File: src/earnings/signal_generator.py (~5 lines, extension of existing 90 %+ LONG skip).

Implementation: Find the existing skip on conviction >= 0.90 and direction == LONG. Add a parallel branch: conviction >= 0.80 and direction == LONG and mcap_tier == "mid" → skip with skip_reason="mid_long_conviction_80".

#4 — Skip SHORTs in non-downtrend

File: src/earnings/signal_generator.py (~5 lines).

Implementation: Replace existing "max 2 SHORTs per session" rule with if direction == SHORT and regime != "downtrend": skip with reason="short_non_downtrend".

#5 — Daily 3-stop session breaker — ✅ IMPLEMENTED May 30, 2026 (Finding #312)

Files (AS SHIPPED, #312): src/earnings/signal_generator.py (_count_todays_stops() helper + skip block in _build_trading_signal after Queue Rule #1) + src/earnings/session_stop_breaker.py (NEW pure-decision module) + src/earnings/config.py (SESSION_STOP_BREAKER_ENABLED / SESSION_STOP_BREAKER_THRESHOLD).

Implementation: In _build_trading_signal (just before final signal construction), query SELECT count(*) FROM trade_audit_log WHERE exit_reason ILIKE '%stop%' AND DATE(exit_time AT TIME ZONE 'America/New_York') = CURRENT_DATE AT TIME ZONE 'America/New_York'. If ≥3, log skip_reason="session_3_stop_breaker" and skip.

⚠️ Note: use exit_reason ILIKE '%stop%', NOT the broken stop_triggered boolean (Finding #233). The latter undercounts badly (~1.7 %) vs the real stop frequency — see the live dashboard / FINDINGS for the current figure.

#6 — BMO + VIX-bucket sizing — 🛑 RETRACTED Apr 25 (Finding #232)

See §4 for the retraction narrative. Briefly: 1.2× scale-UP arm fails ex-top-3 (the +$100 winner bucket collapses to −$1.58 across 17 trades after dropping STT/VZ/ALLY); direct multiplier modeling shows total per-100 lift is +$36 not +$60; operator dropped both arms. Re-test the scale-DOWN arm only at trade #150+.

#7 — RuleB (±90 min stop window) — implementation notes (Phase 1 DRY-RUN shipped May 28, 2026, Finding #305 · 🛑 Phase 2 CANCELLED Jul 5, 2026, Finding #330)

AS-SHIPPED implementation differs from the original spec below. Pre-coding investigation (May 28 afternoon) caught a structural inconsistency in the original spec: it described a "soft gate / bookkeeping defer" approach that would deliver $0 EV (same failure mode as FA_soft retracted in Finding #228). The simulator's simulate_window_only (scripts/analyze_finding_a_full.py:443) requires the hardware stop to actually be SILENT during the ±90min window. Operator chose Path (A) hardware cancel-and-rearm via AskQuestion. See docs/FINDINGS.md Finding #305 for the full as-shipped narrative.

Files (AS SHIPPED): - src/earnings/ruleb_window_gate.py — NEW module with pure decision function decide_ruleb_action() + RuleBAction enum + RuleBDecisionContext dataclass + derive_earnings_release_dt() helper (~180 lines including comprehensive module docstring). - src/earnings/trading_orchestrator.py — _process_ruleb_window_gate() dispatcher + _ruleb_fire_cancel() + _ruleb_fire_rearm() + _ruleb_log_once() helper (~220 lines including detailed comments). - src/earnings/config.py — 4 new env vars: RULEB_ENABLED (default False), RULEB_MODE ('dry_run' / 'active', default 'dry_run'), RULEB_WINDOW_MINUTES (default 90, matches simulator), RULEB_HARD_FALLBACK_MIN_BEFORE_EXIT (default 30, safety boundary). - tests/unit/test_ruleb_window_gate.py — 25 unit tests covering all 8 RuleBAction return values + sanity-gate skips + late-entry edge case + hard-fallback boundary.

Mechanism (AS SHIPPED — hardware cancel-and-rearm, NOT soft-gate): - At t = release_dt - 90min: cancel hardware STP-LMT via IBKRClient.cancel_order(stop_order_id). Stop now silent. - At t = release_dt + 90min: re-place hardware STP-LMT via order_executor._place_stop_order(symbol, shares, side, stop_price, primary_exchange). Stop restored. - Hard fallback: if cancelled state persists past scheduled_exit - 30min, force re-place (catches stuck-cancelled-past-rearm edge case). - State is derived per-cycle from (now, hardware_stop_active, release_dt, scheduled_exit, entry_time) — no DB persistence (memory.md "IBKR is source of truth"). On bot restart, state re-derives from current world.

Failure modes (all handled fail-soft): - IBKR disconnected during cancel/rearm → log error + return; next cycle retries. - Cancel returns False → log warning + retry next cycle (stop still active = position protected; only EV opportunity lost). - Rearm fails → HIGH-severity Telegram alert + retry every cycle (position unprotected until success). - Bot crash between cancel and rearm → on restart, derived state shows "no stop active + past rearm" → re-place fires next cycle. - Restored position with signal=None (Phase 1 limitation) → graceful SKIP_NO_EARNINGS_DT with once-per-position warning log. - Late entry (entry ≥ cancel_target) → graceful SKIP_LATE_ENTRY (stop runs naturally; rule doesn't apply).

Phase 1 → Phase 2 rollout: - Phase 1 (this commit): RULEB_ENABLED=false (production default — operator must explicitly enable). When enabled, RULEB_MODE='dry_run' logs every cancel/rearm action it WOULD take, without IBKR side effects. Operator reviews dry-run logs for ~few days. Phase 2 (separate later session) flips RULEB_MODE='active'. Phase 2 ship is a pure config-flip commit (~1 LOC); may also include the restart-safe earnings_release_dt lookup if Phase 1 logs show that's needed in practice.

Acceptance test (AS RUN): - ✅ pytest tests/unit/test_ruleb_window_gate.py -v → 25/25 PASS pre-commit on prod's Python 3.12 venv. - ✅ Broader regression: pytest tests/unit/test_ruleb_window_gate.py tests/unit/test_signal_regime_gate.py tests/unit/test_signal_volume_threshold.py tests/unit/test_signal_reliability_gate.py → 36/36 PASS, no regressions. - Live post-deploy: enable via RULEB_ENABLED=true env var; monitor journalctl -u earnings-trader | grep "RuleB\[" for dry-run "would CANCEL" / "would RE-PLACE" entries on the next BMO/AMC trade.

#7 (ORIGINAL SPEC, kept for historical context — DO NOT IMPLEMENT) — RuleB (±90 min stop window) soft-gate version

The text below was the original spec WRITTEN ON APR 25 before the as-shipped investigation. It describes a soft-gate approach that would deliver $0 EV. Preserved here only as a historical reference for future agents debugging the "why is the as-shipped implementation different from the spec?" question.

File (ORIGINAL SPEC): src/earnings/order_executor.py + EventBus stop-fill handler (~25 lines).

Implementation (ORIGINAL SPEC — DO NOT IMPLEMENT): On the orderStatusEvent callback for a stop fill, check (now - position.earnings_release_dt).total_seconds() if the release timestamp is known. If abs() < 90 * 60, log "stop blocked: in earnings ±90min window" and DO NOT act on the fill in the bot's state machine (the broker has already filled — this is a bookkeeping defer; the position is reconciled at the next scheduled exit attempt using existing position_manager reconciliation).

Caveat (CORRECTED MAY 28): The original spec claimed "This is the same 'soft gate' pattern as Finding A. It works for RuleB but only because the ±90 min window is narrow enough that the broker's hardware fill + bot's reconciliation-after gives a similar outcome to 'no stop in window.'" This claim was wrong: the broker's hardware fill closes the position at the stop-fired price regardless of window narrowness. The bot's reconciliation doesn't change the trade's P&L. The Finding #228 FA_soft retraction proved this — FA_soft delivered $0 EV because the hardware stop fired anyway. The as-shipped Finding #305 implementation uses hardware cancel-and-rearm to match the simulator's actual mechanism.

7. Deferred questions / data gaps

Tracked from docs/STRATEGY_AUDIT_v3.md §9. Re-test after another ~50 closed trades.

Possible re-test later: H4 (day-of-week / Thursday alpha) — was rejected because Thu's +$209 was the GPRO/DELL/SWBI cluster again. Re-test once the 3-outlier dominance dilutes.

May 28, 2026 LATE evening update — Phase 3 simulator PR 1 COMPLETE. See docs/PHASE_3_SIMULATOR_DESIGN.md §12 for the 5-PR ship plan. Today's session shipped all 5 PR-1 commits:

Commit	Finding	Description
919c566	#307	snapshot_schema='pm_v2' conditional-label fix (closes original Finding #273 Phase 1)
de0d28e	#308	Bug #6 hyp_pnl re-run + structural-ceiling discovery (deadband-conviction rows are un-simulatable)
615befa	#306.a	Phase 3 simulator foundation skeleton (scripts/phase3_simulator/ package + Trade/load_trades duplicated per Q1 HYBRID + RuleSet dataclasses + CLI stub)
06cae0b	#306.b	cost_model.py (proxy-flat-1pct per Q2) + replay.py + reference_adapter.py + cross_check_baseline.py (operator-runnable §10.1 acceptance script)
62cb607	#306.c	gauntlet.py (5 checks per memory §K) + output.py (deterministic JSON schema) + runner.py --run-gauntlet + H-rule self-validation + reproducibility test

All 4 PR 1 acceptance gates PASS (verified on prod): - §10.1 cross-check vs analyze_finding_a_full.py: 100% match within $1 on all 7 rules (A_base/FA_soft/FA_hard/RuleB/RuleD/RuleG/RuleF), max drift $0.00 across 121 trades. - §10.2 reproducibility: 2 back-to-back runs of runner.py --run-gauntlet produced byte-identical JSON output (after masking generated_at). - §10.3 H-rule self-validation: synthetic v2 H-rule shape returns FAIL-EX-TOP-2 per test_h_rule_synthetic_returns_fail_ex_top_2 (would emit DO-NOT-SHIP-PR-2 diagnostic if it ever flipped to PASS). - Phase 0 §0.6 Gate 1/4: corrected per #308 — the signal_emitted 100% subset is sufficient; ~69%/~28% on generate_signal/depth is structural deadband-conviction ceiling, not a coverage gap.

Live gauntlet on prod (runner.py --run-gauntlet --limit 200): RuleB verdict PASS, n=121, per-100 $+90.74 (headline) / $+39.93 (ex-top-2), ship recommendation SHIP. (Numbers are variant-vs-A_base-replay deltas; reference simulator's narrative numbers were variant-vs-LIVE deltas — both correct, see §1 verdict banner.) 🛑 SUPERSEDED Jul 5, 2026 (Finding #330): this PASS/SHIP was an artifact of the pre-#330 gap-fill optimism in simulate_window_only; gap-aware RuleB is FAILED-GAUNTLET (−$194/100 ex-top-2, n=129). Kept for traceability — the §10.1 cross-check itself remains valid (both sides used the same pre-#330 rule math).

PR 2 — Volume-gate retune sweep — ✅ DONE; gate_uniform_$1000 SHIPPED (Finding #311, May 30). (Was "begins next session" — highest-leverage per May-28 PM volume-collapse evidence.) Added gate_variants.py ($2k → $1k → $500 → $250 → $100 → $0 sweep) + replay.py extension for counterfactual admission on skip_reason='low_volume' rows + first phase3_sim_gate_retune dashboard section in strategy_performance_analysis.py + operator AskQuestion ship decisions per variant.

🕓 DEFERRED-POST-PHASE-3 — comprehensive re-evaluation items (parked Apr 28, 2026; ✅ now CLOSED as of May 31, 2026 — Phase 3 simulator PR 1-5 COMPLETE + Phase 4 dashboard consolidation COMPLETE, #306-#322)

✅ May 31, 2026 status — PR 1-5 ALL COMPLETE + Phase 4 dashboard consolidation COMPLETE (#306-#322). (May 28 status, kept for context: PR 1 complete — foundation + replay + cost-model + cross-check + gauntlet + H-rule self-validation + reproducibility + JSON output across 5 commits — see §1 verdict banner.) Final outcomes of the PR 2-5 sequence per docs/PHASE_3_SIMULATOR_DESIGN.md §12: - PR 2 — Volume-gate retune sweep — ✅ DONE; SHIPPED gate_uniform_$1000 (#311). gate_variants.py declarative thresholds + replay.py counterfactual-admission extension for skip_reason='low_volume' rows + first phase3_sim_gate_retune dashboard section. HIGHEST-LEVERAGE per May-28 PM volume-collapse evidence (Jan→May avg cumulative volume dropped ~109×; the then-$2k-gate pass rate had collapsed to 1.9%). Closed the volume-gate retune sub-item of the sizing-priority-v2 row below. - PR 3 — Sizing + priority + exposure-cap variants — ✅ DONE. sizing_policies.py mcap-tier-pct sweep + tier-collapse + per-direction SKIP rule revalidation + position×regime interaction + capital_simulator.py for realistic overlapping-position capital reservation. Outcomes: priority 0-6→0-8 SHIPPED (#313); sizing 30/22/15 KEEP (#314.c — the then-live tiers, since superseded by #314.f/#366); per-cell SKIP rules analysis-only (#314.e); exposure cap — rare/normal/busy collapsed to a single uniform 140% cap (#317 SHIPPED, after #316 showed the rare discount near-inert). Closed the rest of the sizing-priority-v2 row. - PR 4 — Regime classifier v2 candidates — ✅ DONE; KEEP compound_v1 (#318). regime_classifiers.py baseline compound_v1 + EMA20 + MA50 + drawdown OR clause (VIX-term-structure / breadth / credit / FCI remain deferred to Phase 4 nice-to-have). Closed the regime-detection-v2 row below. - PR 5 — Depth-gate v2 decomposition — ✅ DONE; KEEP gate (#319). depth_gate_variants.py decomposed spread_gate + depth_gate variants + threshold sweep. Closed the depth-gate v2 row below.

Each PR: ship code + integrate to runner.py + add phase3_sim_* dashboard section + run on prod + operator AskQuestion ship decision per variant before any production rule change.

Original framing (Apr 28, 2026 — kept for context): the bundle items were parked because doing them before the Phase 3 simulator landed would conflate rule-effects with infrastructure-effects and retroactively shrink the live-evidence buckets the rules are sized on. Mirror in dashboard 🎯 Action Queue under status=DEFERRED-POST-PHASE-3 (build_action_queue() entries regime-detection-v2 + sizing-priority-v2); these have now all flipped to CLOSED/IMPLEMENTED as PR 2-5 shipped (#311-#319) per design doc §12.

Saved artifacts (per memory §J/§K/§N — both items have: canonical Finding in docs/FINDINGS.md, forward entry here, AGENT_HANDOFF §11 DEFER bullet, programmatic Action Queue row in build_action_queue(), and dashboard render under DEFERRED-POST-PHASE-3 status filter). Re-runnable analysis script for the regime item is committed at scripts/analyze_regime_detection_lag.py.

8. Cross-references

Scripts (rerunnable)

Documentation

Dashboard

URL: http://[hidden]:8080/#strategy

This playbook is the authoritative forward-looking reference. Update it when a rule ships (move from §2 queue → §3 verified), when a rule fails (move to §4 retracted), or when a deferred question becomes testable (move from §7 to §3 if validated).