Strategy ROI Audit v3 — Verified Findings + Audit Trail

🛑 Jul 5, 2026 update — RuleB ALSO RETRACTED (see §16; Finding #330)

The Apr-25 banner below said "Only RuleB survives". That is no longer true. The Jul 5 strategy review found simulate_window_only filled window-blocked stops at the STOP PRICE after the price had gapped through — unobtainable under the as-shipped #305 cancel-and-rearm mechanism. Gap-aware (rearm fills at the first post-window bar OPEN, the same rearm_gap_open handling that retracted FA_hard): RuleB = −$209/100 headline / −$194/100 ex-top-2 (n=129) — FAILED-GAUNTLET; RuleG also flips negative. NO stop-loss rule variant currently survives. RuleB Phase-2 active flip CANCELLED; v3 D / v3 E lose their RuleB dependency (re-evaluate as pure selection rules). Full record: §16 + canonical narrative docs/FINDINGS.md Finding #330. All RuleB-survives statements in §§4-14 below are the historical pre-#330 record.

---

⚠️ Apr 25, 2026 update — Finding A RETRACTED (see §10)

The comprehensive Finding-A simulator (scripts/analyze_finding_a_full.py, Apr 25) showed that v3 Finding A — Block stops outside RTH is either inert (rule as written, software-only gate while hardware STP-LMT stays outsideRth=True per Finding #179) or net-negative (hardware also gated: −$418 / 100 trades, opposite sign to the audit's claim). Findings D / G also fail the 4-check bias gauntlet on the simulator's apples-to-apples baseline. Only RuleB (±90 min window) survives, at a much smaller +$123 / 100 trades (🛑 superseded Jul 5, 2026 — see the banner above). Full retraction + replacement queue in §10. Canonical narrative: docs/FINDINGS.md Finding #228. New forward-looking source of truth (this doc + dashboard Strategy Analysis tab + simulator data combined): docs/STRATEGY_PLAYBOOK.md.

---

Purpose: Permanent record of the v3 audit pass (Apr 22, 2026) so future analyses can see what was tested, what survived bias-checks, and what was retracted. Mirrors the canvas at ~/.cursor/projects/Users-marius-projects-polymarket-bot/canvases/strategy-roi-audit.canvas.tsx (Cursor-local, not in git).

Read order: §1 process learning → §2 v2 retractions → §3 v3 verified findings → §4 ranked recommendations → §5 audit trail (every hypothesis tested) → §6 statistical caveats → §10 Finding A retraction (Apr 25).

1. Process learning — the H-rule retraction

In v2 the "ATR > 5% pre-screen" rule looked like a clean +$329 winner when measured on the 49 stopped-correct trades alone. The full-sample re-check killed it: across all 109 trades the same rule would have removed +$163 of profit (mostly SHORT alpha) and turned the strategy net −$148 worse.

The diagnostic showed something deeper: 3 trades — GPRO (+$111), DELL (+$80), SWBI (+$70) — together carry ~$260 of the strategy's $61 net P&L. Multiple v2 "findings" (VIX 22+ is best, L 70-80% is the only profitable LONG band, Thursday is the alpha day, Mid-cap-SHORT bucket) all decompose into "the same 3 outliers happen to land in this slice." Without those 3 winners, the strategy is roughly −$200.

The 4-check gauntlet (now mandatory for every new finding)

1. Raw signal + n adequate — n ≥ 12 minimum, ideally n ≥ 20.
2. Drop top-2 outliers — does the finding survive without them?
3. De-dupe by unique ticker — does the finding survive when repeated evaluations of the same ticker are counted once?
4. Full-population spillover — does applying the rule break un-targeted trades? (Especially critical for skip rules.)

Anything that fails any check is rejected regardless of how good the isolated number looks.

The H-rule full-impact table (the smoking gun)

Run via scripts/analyze_atr_prescreen.py against all 109 closed trades:

The skipped trades carry the strategy's profits, not its losses. By-direction at the 5 % cutoff:

By timing at 5 %: AMC: skip 24/49 (49 %), skip_pnl = +$126 — would amputate half the AMC book including the names that drive AMC SHORT alpha.

Lesson: had we shipped rule H based on the isolated stopped-correct slice (where it showed +$329 saved), the strategy would have gone net −$148 across the full population.

2. v2 findings — bias re-check verdict

3. v3 verified new findings (all post-bias-check)

Finding A — Block all stop-loss orders outside RTH (09:30–16:00 ET)

🛑 RETRACTED Apr 25, 2026 — comprehensive simulator showed the rule as written is inert (software gate; hardware STP-LMT stays outsideRth=True) and the only effective form (outsideRth=False extension) is −$418 / 100 trades, not the claimed +$300–$400 / 100 trades. The audit treated Rule B as a proxy and conflated the per-session observation table (which measures where stops fired, not what would have happened if they hadn't) with counterfactual evidence. See §10 below for the simulator's full output and docs/FINDINGS.md Finding #228 for the canonical retraction narrative.

Mechanism (as originally claimed): Pre-market and after-hours liquidity is thin; STP-LMT fills at bad prices then stocks revert during RTH.

Evidence: | Session | n | Stops | Total $ | |---|---:|---:|---:| | Pre-mkt 04:00–09:30 ET | 22 | 16 (73%) | −$242 | | RTH 09:30–16:00 ET | 63 | 18 (29%) | +$428 | | After-mkt 16:00–24:00 ET | 25 | 16 (64%) | −$138 |

Bias check: No outlier dominance — 32 stops across 47 ext-hours exits are spread across distinct tickers. Mechanism is structural (liquidity) not selection. Full-population spillover is positive (RTH cell is +$428).

Implementation: Gate stop-trigger check on session-of-now in order_executor; hardware STP-LMT order can stay placed; software gate just refuses to act on triggers outside RTH. ~15 lines.

Estimated lift: +$300 to +$400 / 100 trades.

Stop-counterfactual rule comparison (per-trade simulation, n=49 stopped+correct)

Full simulation results from scripts/analyze_stop_counterfactual.py --population correct (Polygon 1m bars, exact replay vs alternative stop-loss policies). Live total P&L on these 49 trades: −$584.

~~H~~ retracted per §1 — the +$329 was a selection-bias artifact; full- population check showed −$148 net. Surviving candidates ranked by risk-adjusted lift: B is the safest (only 3 trades worse than live, no selection change, simplest implementation); D adds extra savings at the cost of 19 still-stopped trades exiting at wider losses; G is the combined optimum.

The chosen v3 Finding A is a stronger version of B — instead of just a ±90-min window around the earnings release, it blocks ALL stops outside RTH (09:30–16:00 ET). Rationale: 80 % of stop bleed (−$380 of −$549) came from extended-hours exits regardless of how close to the release the stop fired. Per-session detail in the table at the top of this section.

Wrong-prediction validation (does the rule still catch disasters?)

Run via scripts/analyze_stop_counterfactual.py --population wrong (n=16 trades where prediction_correct=FALSE):

Methodological caveat: the simulator does NOT model the bot's instant_exit:earnings_miss path (the existing earnings-detector that catches most wrong predictions ≤ 60 s after release). 12 of 16 live wrong-prediction exits used that path, not the stop. So the wrong-pop $ deltas are upper bounds. The important conclusion is the direction: every alt-stop rule is positive on the wrong-prediction set, meaning none of them remove the strategy's ability to catch genuine disasters. The existing earnings_detector does the heavy lifting on wrong calls; the stop-loss is the safety net behind it.

Per-trade detail: scripts/stop_counterfactual_correct.csv (49 rows, 35 columns) and scripts/stop_counterfactual_wrong.csv. Columns include A_baseline_*, B_window_*, C_atr15_*, D_atr20_*, E_atr25_*, F_nostop_*, G_combo_*, H_*, atr_daily, atr_pct_of_entry. Sort by {rule}_pnl_d - live_pnl_dollars to inspect winners vs losers per rule.

Finding B — Skip LONGs in downtrend regime (kept from v2 #1, validated)

Evidence: 22 LONG trades in downtrend: 5 wins, −$136, 14 stops (64% stop rate). Same ticker universe in uptrend: 34 trades, 15 wins, +$71, 32% stop rate. Half the win rate, double the stops.

Bias check: 17 of 22 negative; ex-worst-2 (BABA −$32, GME −$25) still −$80; top losers are 5 distinct tickers (no single name dominance).

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

Estimated lift: +$140 / 100 trades.

Finding C — Skip LONGs with pre-event drift ≥ 2% from prior-day close

Mechanism: When a stock has already drifted into our LONG direction overnight, the EPS expectation is "priced in" — the small additional move on the announcement isn't enough to cover commission + slippage + stop risk.

Evidence: | Drift bucket | n | Wins | Total $ | |---|---:|---:|---:| | LONG drift < 2% | 75 | 31 (41%) | +$107 | | LONG drift 2–5% | 13 | 1 | −$120 | | LONG drift ≥ 5% | 2 | 0 | −$25 |

Bias check: 14 of 15 are losers; only EBAY (+$5.64) is positive. Ex-best-1 the bucket is −$150 across 14 losers. Monotonic — the more drift, the worse the outcome. No outlier rescue.

Implementation: New skip_reason='pre_event_drift_long' in signal_generator; uses earnings_markets.stock_price_day_before_close already populated by Finding #227 resolver (713/715 past markets covered). ~15 lines.

Estimated lift: +$145 / 100 trades.

Finding D — Skip Mid-cap LONGs at conviction ≥ 80%

Evidence: | mcap × conviction | n | Wins | Total $ | |---|---:|---:|---:| | Mid LONG 80–90% | 31 | 11 (35%) | −$56 | | Mid LONG 90–95% | 15 | 1 (7%) | −$55 | | Large LONG 80–90% | 9 | 4 (44%) | +$22 |

Bias check: Both Mid sub-buckets independently negative across n ≥ 15. Mid is 70% of our trade count, so this is a high-leverage rule. Large 80-90% is positive — keep that bucket; only Mid is the problem.

Implementation: Extension of v2 finding #5 (90%+ LONG skip across all mcaps) lowered to 80% for Mid only. ~5 lines added to existing skip logic.

Estimated lift: +$110 / 100 trades.

Finding E — Skip SHORTs in non-downtrend regimes (replaces v2 #3)

Evidence: SHORT in uptrend: 1/7 wins, −$2. SHORT in sideways: 0/5 wins, −$27. Combined non-downtrend: 0 wins out of 12 (excluding 1 marginal $0.47 outlier), −$63. SHORT in downtrend: 3/3 wins (GPRO-dominated +$173 but ALL 3 winners across distinct tickers).

Bias check: 12/12 negative on the skip side — perfect direction signal. The downtrend SHORT positive side is dollar-dominated by GPRO, so the estimated dollar lift on the skip rule is conservative.

Implementation: Conditional gate: skip SHORT when regime != 'downtrend' → skip_reason='short_non_downtrend'. Replaces existing "max 2 SHORTs" rule (which was over-cautious).

Estimated lift: +$60 / 100 trades + lower tail risk.

Finding F — BMO + VIX-bucket sizing

Evidence: | BMO × VIX | n | Wins | Stop rate | Total $ | |---|---:|---:|---:|---:| | BMO + VIX < 18 | 18 | 11 (61%) | 17% | +$100 | | BMO + VIX 18–22 | 30 | 11 (37%) | 53% | −$54 | | BMO + VIX 22+ | 6 | 1 | 17% | +$19 |

Bias check: BMO+VIX<18 cell ex-top-3 (STT, VZ, ALLY) still +$32 across 15 trades. Loser cell ex-top-3 (WING, BLK, GRMN) is −$160 across 27 trades — LOSERS are pervasive, wins concentrated. Both ends robust.

Implementation: risk_manager weighted multiplier on VIX bucket for BMO trades only: VIX < 18 → 1.2× base; VIX 18–22 → 0.6×; VIX ≥ 22 → 1.0×. AMC unchanged (no clean VIX signal there). ~10 lines.

Estimated lift: +$60 / 100 trades.

4. Ranked recommendations (revised Apr 25, 2026 — post-simulator)

⚠️ Original §4 ranking was built on the un-validated v3 Finding A and on the pre-simulator estimates for D / G. The Apr 25 simulator killed Finding A and showed D/G fail the 4-check gauntlet. The new ranked queue below is the operative one. Original rows kept inline (struck through) for audit history.

Combined realistic lift (post-retraction): Skip rules #1+#2+#3+#4 overlap (all subset-of-LONGs); realistic combined lift is +$300–$450 / 100 trades. RuleB adds a marginal +$50–80 net of overlap. Total addressable improvement is +$350–$500 / 100 trades, not the +$500–700 the original v3 §4 claimed (the difference is the retracted Finding A and the gauntlet-failed D/G).

The combined queue still flips the strategy from −$60 / 100 trades (current) to roughly +$300 to +$440 / 100 trades, but via selection fixes (cut bad trades upstream) rather than execution fixes (try to rescue already-bad trades).

Implementation order (revised): Ship in this order so each rule's lift is attributable:

1. Skip LONGs in downtrend (Finding B) — single SQL gate in signal_generator, ~10 lines, uses existing MarketRegimeProvider.
2. Skip LONGs with drift ≥ 2 % (Finding C) — uses already-populated earnings_markets.stock_price_day_before_close, ~15 lines.
3. Skip Mid-cap LONGs at conviction ≥ 80 % (Finding D) — extension of existing 90 % skip, ~5 lines.
4. Skip SHORTs in non-downtrend (Finding E) — replaces over-cautious "max 2 SHORTs" rule.
5. Daily 3-stop session breaker — tail-risk circuit breaker.
6. VIX-bucket sizing for BMO (Finding F) — sizing layer.
7. RuleB ±90 min window — only after selection rules settle and we want a marginal stop-loss tweak.

Re-run scripts/strategy_performance_analysis.py and scripts/analyze_finding_a_full.py after each deploy. Re-validate the entire audit at trade #150+ via the 4-check gauntlet (memory §K).

5. Audit trail — every hypothesis tested

6. Audit-process learnings (for next pass)

1. Always run the full-population check before recommending a skip rule. The H-rule (ATR > 5% pre-screen) showed +$329 on the 49-trade subset and would have lost $148 net across 109 trades. Skip rules in particular need full-population spillover analysis because they remove WINNERS as well as losers.

2. The 3-outlier curse. GPRO + DELL + SWBI together dominate ~$260 of the strategy's $61 net P&L. ANY slice that includes them looks great; ANY slice that excludes them is structurally negative. Always report numbers ex-top-2 alongside the headline.

3. Repeated-ticker bias. The MBWM finding in v2 was 9 evaluations of 1 ticker presented as "9 wins." Every counterfactual table needs a unique_tickers column alongside n_total.

4. Mechanism beats correlation. The findings that survived (extended-hours stops, drift filter, BMO + low VIX) all have a structural mechanism behind the data (liquidity gaps, price-priced-in, low pre-earnings vol). The retracted ones (VIX 22+ best, Thursday alpha) were correlation-only and traced back to the same outlier names.

5. The defer queue is a feature, not a failure. 11 of 24 hypotheses ended up DEFERRED. They're not bad ideas; they need either better data capture (H8 conviction_at_exit, H19 pre-event after-hours quote) or larger n (H10 same-ticker, H23 earnings-detector latency). Re-run them in 50 trades.

7. Statistical caveats

n=109 closed trades + n=126 calibrated markets means ±5–8% on headline win-rate slices, ±10–15% on regime-conditional and VIX buckets. Findings A, B, C, D are well-sampled (n ≥ 15 each) and robust to outlier removal. Finding E (skip non-downtrend SHORT) is 12/12 negative — perfect direction but small dollar impact. Findings F (BMO + VIX < 18) and the kept v2 #7 (90%+ LONG skip) are well-sampled.

Re-validate the entire audit after the next 50 trades with scripts/strategy_performance_analysis.py per memory §K. The bias-check gauntlet (4 checks per hypothesis) should be re-run every audit pass.

8. Cross-references — full artifact catalog

Canvas (interactive, Cursor-only)

• ~/.cursor/projects/Users-marius-projects-polymarket-bot/canvases/strategy-roi-audit.canvas.tsx — mirror of this doc with charts, KPI strip, and per-finding deep-dive cards.

Audit scripts (rerunnable)

Live data infrastructure (Finding #227)

Documentation

Operational notes captured elsewhere

• Polygon Stocks Starter ($29/mo) confirmation — burst-tested 50 req/5 s with 0 × 429s. Default --polygon-sleep 0.05. Documented in .env.example and docs/OPERATIONS.md.
• Polygon EPS coverage progression — actual_beat coverage on past earnings_markets rows (n=715):
• Initial (Yahoo + strict 0.99/0.01 PM proxy): 179 / 715 = 25 %
• • Polygon /vX/reference/financials (Tier 3 EPS): 366 / 715 = 51 %
• • Polymarket cascade (Tier 4: post-event yes_price + relaxed thresholds): 524 / 715 = 73 %
• Remaining 27 % gap: markets with no EPS source AND no Polymarket conviction snapshot (mostly pre-conviction-tracker condition_ids).
• Update (#321): current actual_beat coverage is ~95-99.8% after the post-event snapshot + SEC EDGAR + resolver-guard fixes (#246/#247/#321). The 73% above is the Apr-25 audit-trail figure, retained for the progression history.

9. Open questions / DEFER queue

11 hypotheses ended up DEFERRED rather than rejected — they're not bad ideas, they need either better data capture or a larger n. Re-test in the next audit pass (after another 50 trades, ~Aug 2026).

One rejection that may want re-test: H4 (day-of-week) was rejected because Thu's +$209 was the same GPRO/DELL/SWBI outliers. Once we have 50+ more trades (and the 3-outlier dominance dilutes), re-run H4 — if Thursday is still positive, it might be a real session effect not just outlier coincidence.

10. Finding A retraction — comprehensive simulator (Apr 25, 2026)

Why this section exists: The original v3 §3 Finding A claim (block stops outside RTH = +$300–$400 / 100 trades) was supported by (a) Rule B as a proxy and (b) a per-session observational table. A follow-up review surfaced 8 unmodelled scenarios. The script scripts/analyze_finding_a_full.py was built to close all 8 and replay the actual proposed rule end-to-end.

10.1 The simulator — what it does

scripts/analyze_finding_a_full.py (570 lines, read-only):

• Replays every closed trade against Polygon 1m bars.
• Distinguishes two versions of Finding A:
• FA_soft — software-only gate; hardware STP-LMT keeps outsideRth=True per Finding #179, so the broker still fills in extended hours regardless of any software gate. By construction the price evolution is identical to current behavior.
• FA_hard — hardware also flipped to outsideRth=False outside RTH; stop is silent in pre/post; on first RTH bar after a non-RTH segment, if the bar gap-throughs the stop (open ≤ stop for LONG / open ≥ stop for SHORT) we exit at the open price (gap-open slippage modeling), otherwise the stop re-arms and intraday triggers fire as normal.
• Models earnings_detector cap for wrong-prediction trades (the recorded exit_time is treated as a hard ceiling — the simulator can't credit the rule for trades the live instant_exit:earnings_miss pipeline already catches).
• Halt-suspect detection (>30 min bar gap + >3 % price jump in extended hours).
• Cross-compares against Rules B / D / G / F on the same fetched bars.
• Built-in 4-check bias gauntlet per rule: headline / ex-top-2 outliers / ex-repeat-tickers / by prediction-correctness / by FA_hard action / pre- EPS-trigger only.

10.2 Results (population = stopped-only, n = 51 trades)

Apples-to-apples vs the simulator's A_base baseline (cancels Polygon-bar- vs-IBKR-fill reconstruction noise):

10.3 Mechanism — why FA_hard loses money

The 22 worst trades include MDB (−$82), TTD (−$58), NKE (−$54), BABA (−$38), WDAY (−$26), NYT (−$23), RL (−$18), GTLB (−$15), HOOD (−$15), LOW (−$13), CSCO (−$12), VSCO (−$12). Every one of these had the stock continue past the stop level into RTH open. The audit's claim that extended-hours stops "fill at bad prices then stocks revert" assumes reversion that the data does not support — for these 22 trades the price moved the wrong direction and stayed there.

The 9 recovery wins (KSS +$66, GME +$33, BBY +$31, KR +$23, JPM +$15, TXRH +$11, WMT +$11, NET +$10, SOUN +$4) are real but outweighed 2-to-1 by the gap-open losses in both trade count and dollars.

10.4 Bias gauntlet — D/G also fail

The audit's "best" stop-loss alternative (RuleG: window + ATR k=2.0) showed +$200 / 100 headline. Robustness:

Both collapse to net-negative when you drop just two outliers (the same KSS / GME / BBY recovery cluster). They fail bias-check #2 of the gauntlet → rejected as standalone rules.

10.5 What survives — RuleB only

Rule B (±90 min no-stop window around earnings release):

• +$123 / 100 trades vs A_base.
• 4 better / 1 worse / 51 unchanged → 96 % no-effect.
• Robust to ex-top-2 (still positive at smaller magnitude).
• Mechanism: disables the stop only inside the ±90-min release window, which is the period when earnings volatility legitimately spikes the stock through stop levels even when the prediction is correct.
• Risk profile: very low — the rule only changes 5 of 51 stopped trades.

This is the only stop-loss intervention worth shipping, and only after the selection rules (Findings B/C/D) have been deployed and measured.

10.6 Pre-EPS-trigger only (the audit's strongest claimed scenario)

The audit claimed extended-hours stops fire "on overnight noise before any earnings." Restricting the simulator to the 30 trades whose A_base trigger fired before the earnings release: FA_hard = −$92, 11 better / 14 worse. Even in this allegedly safest subset, the rule is net-negative. The "noise" hypothesis isn't supported — pre-EPS extended-hours stops fire on real adverse moves that continue through RTH open more often than they reverse.

10.7 Halt-suspect cases caught by the simulator

Two trades flagged by the halt heuristic (>30 min gap + >3 % jump in pre/ post-market):

• ARES (LONG BMO): 134-min gap, +3.71 % jump on Feb 5 between 05:17 and 07:31 ET.
• NYT (LONG BMO): 41-min gap, +3.92 % jump on Feb 4 between 07:05 and 07:46 ET.

Under FA_hard the bot would have been forced to hold through both, with no ability to act on the price gap until RTH. These are real halt risks not modelled in the audit.

10.8 Cross-references

• Simulator script: scripts/analyze_finding_a_full.py
• Per-trade output: scripts/finding_a_full.csv (51 rows × 41 cols)
• Canonical retraction narrative: docs/FINDINGS.md Finding #228
• Forward-looking source of truth combining v3 + simulator + dashboard state: docs/STRATEGY_PLAYBOOK.md

11. Universe View follow-up (Apr 25, 2026 — afternoon refresh)

A second analysis pass extended the audit to the full universe of all 900 Polymarket earnings markets in the database, not just the 109 the bot actually traded. The output is rendered live in the dashboard 🌐 Universe View section (refreshed every report run) and also mirrored in the universe_hit_rate block of scripts/strategy_analysis_report.json.

11.1 What it confirmed

• The conviction signal IS profitable at universe scale. 253 LONG candidates with prob_24h_before > 0.60: dayClose win rate 55.3 %, $10K-notional sum +$17,831. 26 SHORT candidates with prob_24h_before < 0.30: dayClose win rate 61.5 %, sum +$3,124. Polymarket's own prediction accuracy is 75 % LONG / 88 % SHORT.
• The 90-95 % LONG bucket is independently bad at universe scale (50.9 % WR, −$2,886 across 55 markets). Independently corroborates the Mar 9 implementation (Finding #202) that skips this band — two lenses, same conclusion, 4× larger sample on the universe side.

11.2 What it complicated

v3 Finding C (drift filter, Queue #2) — verdict: WEAKENED. The original v3 C drift definition (live_entry_price − stock_price_day_before_close) collapses to ~0 for BMO trades because the bot enters at 15:58 ET and stock_price_day_before_close is the same day's 16:00 close. Replacing it with a 5-trading-day baseline (universe view § C) and re-bucketing the 253 LONG candidates:

Skip-bucket (drift ≥ 2%): n=74, win 55.4 %, total +$3,289, ex-top-2 +$1,136. Skipping these would remove winners, not losers. Live trades with the original drift defn still show drift ≥ 2 % is bad (−$120 on n=13), but the sample is at the n ≥ 12 floor and the rule's effect may trace to a different underlying signal (small-caps, low-conviction overlap, etc.) than literal pre-event price drift.

Action: Queue #2 status changed from "IN-QUEUE — clean" to "IN-QUEUE — drift-baseline question must be resolved before shipping." Three options:

1. Use a shorter baseline (e.g. 1 trading day prior) — closer to the v3 live-trade definition; may reproduce the live signal.
2. Use the 5-day baseline but tighten the skip threshold (e.g. ≥ 5 % only, where the universe shows −$315 across 24 markets).
3. Drop Finding C and rely on Findings B + D + E to catch the same trades via different upstream filters.

v3 Finding D (Mid-cap 80%+ skip, Queue #3) — verdict: CONTESTED. Universe view § B splits the LONG 80-90 % bucket by market_cap_category:

The Mid bucket is net positive at universe scale, contradicting the live-trade Mid 80-90 % LONG = −$56 / 31 trades observation. The most likely explanation per Finding #228: stops eat the winners on live Mid-cap trades (held-correct = +5 % / trade vs stopped-correct = −3.5 % / trade). The Mid bucket may NOT need a skip rule — it may need better stop handling (covered by Queue #7 RuleB).

Action: Queue #3 status flagged "CONTESTED" — re-evaluate after RuleB ships. If RuleB closes the gap on Mid-cap stops, Finding D may no longer be needed.

11.3 What it didn't change

• Finding #228 retraction (v3 A) — independent of universe view.
• v3 Findings B (downtrend LONG skip), E (non-downtrend SHORT skip), F (BMO + VIX sizing) — universe view doesn't have regime / VIX columns populated densely enough to re-check these. Live-data evidence stands.
• The 90 %+ LONG skip (Mar 9, IMPLEMENTED) — universe view independently confirms.

11.4 Cross-references

• Function: scripts/strategy_performance_analysis.py:analyze_universe_hit_rate()
• Helper: scripts/strategy_performance_analysis.py:_fetch_5d_prior_close()
• JSON key: strategy_analysis_report.json["universe_hit_rate"]
• Dashboard section: 🌐 Universe View (renderUniverseView in dashboard.html)
• Action queue rows updated: v3 C (Queue #2), v3 D (Queue #3) — see STRATEGY_PLAYBOOK.md §2 with the new "⚠️" markers.

12. Apr 25 evening — operator decisions + volume-aware Universe View

After §11 was written, three additional changes landed in one pass.

12.1 v3 Finding C — RETRACTED (option 3 chosen)

Operator decision: drop the drift filter entirely and rely on B + E + the already-IMPLEMENTED selection rules to catch high-drift bad-EV trades via different filters. Reasoning:

• Universe view at 5-day baseline showed the supposed "skip bucket" is +$3,289 net at $10K notional (gauntlet check: WEAKENED).
• At the bot's $2K volume gate the skip bucket is +$4,195 net (n=27, win 66.7 %) — even more positive when properly filtered. Gauntlet verdict still WEAKENED.
• Live n=13 in the skip bucket is at the n ≥ 12 gauntlet floor — too small to override the universe view's 27-market sample.
• Original v3 C drift definition (live entry_price − stock_price_day_before_close) collapses to ~0 for BMO trades because both are within 2 minutes of each other.
• Three resolution options were considered (1: 1-day baseline ≈ live defn, 2: 5-day with ≥ 5 % threshold, 3: drop entirely). Option 3 was chosen because the live signal isn't strong enough to override the universe weakening, and B + E + the existing 90 %+ skip catch most high-drift bad-EV trades anyway.

Re-test trigger: trade #150+ with a tighter drift definition (e.g. 1-day-prior baseline that approximates the original v3 defn).

12.2 v3 Finding D — ON-HOLD (pending RuleB)

New ON-HOLD status added to the Action Queue schema. v3 Finding D moved there because:

• Universe Mid LONG 80-90 % (unfiltered, n=13) was +$1,175 — opposite sign from live (−$56 across 31 trades).
• At the bot's $2K gate the picture is mixed: Mid n=10 = +$188; Large n=28 = −$382 (the v3 audit's "Large 80-90% KEEP +$22" claim does not reproduce when properly volume-filtered). Neither bucket is cleanly positive at universe scale with the proper gate.
• Most likely cause per Finding #228 mechanism: stops eat winners on live Mid-cap trades (held-correct = +5 % / trade vs stopped-correct = −3.5 % / trade). If RuleB closes the live Mid-cap stop bleed, this rule is no longer needed.

Re-evaluation criterion: after RuleB ships AND has accumulated ≥ 10 Mid-cap stops in production, re-measure live Mid-cap stop rate and P&L.

12.3 Universe View v3 — volume-aware

analyze_universe_hit_rate() extended to:

• Join markets.volume_total for volume-coverage stats. Caveat: only 35 % of in-band markets (111 / 314) have positive volume_total — the rest may have Gamma volume that the bot's effective gate (Finding #224) would still pass, so this analysis is conservative and may UNDER-count the bot's true tradeable universe.
• New "volume sensitivity sweep" sub-section (§ A in dashboard) showing LONG conv > 60 % at $0 / $100 / $500 / $2K (BOT GATE) / $5K / $10K / $25K. Headline finding: unfiltered +$17,831 collapses to +$4,497 at the bot's actual gate (+$42/trade avg vs +$70 unfiltered). Signal remains positive but much smaller.
• 80-90 % mcap split now rendered side-by-side (unfiltered vs $2K gate). Queue v3 D bias_check now reads from the gated split.
• Finding C tighter drift cross-check now computes both unfiltered and $2K-gated buckets + skip-bucket summaries per gate.

12.4 New ranked queue (post Apr 25 evening — first revision)

Combined realistic lift: +$200 to +$350 / 100 trades.

12.5 v3 E ON-HOLD (later that evening — sample-size pushback)

Operator review of v3 E found the rule too aggressive given current data:

• KEEP set is dangerously small. n=3 downtrend SHORTs is below any reliable threshold; the +$172 alpha is dominated by a single GPRO trade (the 3-outlier rule — memory §K — applies). Ex-GPRO it's 2 trades / ~$60.
• SKIP set is essentially flat at headline. 15 non-downtrend SHORTs sum to −$29 gross; the +$58/100 estimate is bias-checked ex-top-2 (which removes STT +$37 and BX +$31). Without those two non-downtrend SHORTs being treated as outliers, the rule barely pays.
• Sideways SHORT 67 % stop rate is suspicious. Same pattern as Mid- cap LONGs (which drove v3 D to ON-HOLD): may be Finding #228 stops-eat-winners on extended-hours noise rather than a selection issue. RuleB will tell us.
• Universe view can't validate. Only 5 SHORT markets pass the bot's $2K volume gate; we can't independently corroborate the regime split.
• Bot is capital-constrained. ~$2,440 trading capital, max 3-5 positions per cycle. SHORT signals are already only 16 % of book. Skipping ~12 future SHORT trades to "save" what may be sideways stop bleed could starve the bot of valid alpha for months between downtrend clusters.

Re-evaluation criteria (strictest of): - After RuleB ships AND total SHORT count ≥ 30 (currently 18); OR - At trade #150+ regardless (audit re-validation gauntlet).

If at re-evaluation: - Sideways SHORT stop rate drops post-RuleB → close as RESOLVED-BY-OTHER-RULE. - Sideways SHORT remains 0 % WR with high stop rate → resume IN-QUEUE and ship. - Downtrend SHORT alpha is still GPRO-dominated at n ≥ 6 → keep ON-HOLD permanently (rule has no reliable KEEP set).

12.6 Final ranked queue (post v3 E ON-HOLD)

Combined realistic lift: +$160 to +$280 / 100 trades (down from +$200-350). Still flips the strategy from −$60 / 100 trades to +$100 to +$220 / 100 trades — driven primarily by Finding B (downtrend LONG skip) plus RuleB and behavioral guards.

Net effect of being conservative on D + E: ~$100 / 100 trades less estimated lift, but we preserve trade volume and avoid committing rules to small-sample data that may be capturing something temporary (capital constraints during a particular regime, the 3-outlier curse, or stops bleed that RuleB will fix).

13. Apr 25 evening — data-quality backfill (Findings #229 + #230)

After the queue updates of §12, two data-quality issues were uncovered and fixed in succession. They don't change any queue rule's status (the ranking from §12.6 is unchanged), but they materially expand the data underpinning every "is this signal valid?" check that depends on the Universe View. Future audits should rely on the post-backfill numbers.

13.1 Finding #229 — markets.volume_total silently 0 for resolved markets

Trigger: Universe View showed only 35 % volume coverage (111 of 314 in-band markets). User asked whether this was a real Polymarket gap or a data-pipeline gap.

Root cause: continuous_collector.sync_markets_from_gamma queried Gamma with ?active=true&archived=false. Polymarket's state machine has both active and closed flags — settled earnings markets sit in (active=True, closed=True, archived=False) for days/weeks before being archived. The ?active=true filter does NOT include them despite their active=true flag.

Fix: two-pass sync (active + closed-not-archived) in continuous_collector + one-off scripts/backfill_markets_volume.py. Recovered 319 markets / $5.7M of volume.

Result: markets passing the bot's $2K gate jumped 111 → 546.

13.2 Finding #230 — earnings_markets.prob_*_before missing for 401 of 900 markets

Trigger: Even after Finding #229, the Universe View was only capturing 31 % of all earnings markets (279/900). The biggest gap was 549 rows with NULL prob_24h_before (401 past + 148 future).

Root cause: Two-layer issue. (1) Orderbook collector (continuous_collector) only starts sampling a market when its condition_id enters the active-markets list, often only a few hours before the earnings event. So the 24h / 12h / 6h-before snapshot windows fall outside the sampled timespan. (2) conviction_tracker --backfill looks for snapshots within ±30 min of the target time — useless when the collector has no rows from that timeframe at all.

Fix: new scripts/backfill_conviction_from_trades.py pulls trade history from Polymarket's data-api /trades endpoint (which goes back much further), normalises each trade's price to YES-token probability (yes_prob = price if trade on YES, 1 − price if on NO), and finds the trade closest to each target timestamp (24h/12h/6h/1h/30min before earnings_dt) within ±60 min.

Result:

13.3 Sign-flips & re-validations

13.4 Action queue impact

Finding #230 strengthened the existing verdicts but did not change any queue ranking or status:

• v3 D ON-HOLD verdict reinforced (Mid 80-90 % at $2K gate is now n=73 at +$381 — still weakest LONG, but still POSITIVE without stops; the Finding #228 stops-eat-winners interpretation is now backed by 5× more data).
• v3 C RETRACTED verdict reinforced (with 529 LONG markets, the drift filter would skip ~150 net-positive markets).
• v3 E ON-HOLD unchanged (SHORT data is still thin, and the 20-30 % bucket sign-flip is a warning against committing to small-sample SHORT findings).

13.5 Forward-fix gap → Finding #231

A new finding (#231) closes the data-pipeline gap so this doesn't re-accumulate:

• Integrate the data-api /trades fallback INTO conviction_tracker.py:backfill_historical_convictions() so the tracker auto-falls-back when price_snapshots is sparse.
• Add a weekly systemd timer (polymarket-data-backfill.timer) that runs both volume and conviction backfills, so the gap self-heals.

See docs/FINDINGS.md Finding #231 for the canonical narrative.

13.6 Data hygiene rule going forward

Single source of truth. All strategy analyses must use the live production tables (earnings_markets, markets, trade_audit_log, polymarket_opportunity_log, skipped_signals, conviction_history, price_snapshots). Derived artifacts (scripts/strategy_analysis_report.json, scripts/finding_a_full.csv) are regenerated from these tables and should never be the source. When data quality issues are uncovered, the fix flow is: (a) write a backfill script to update the production table, (b) write a forward-fix to the collector / tracker so the gap doesn't re-accumulate, (c) re-run the master report so the dashboard reflects the new state, (d) document as a numbered finding in docs/FINDINGS.md.

14. Apr 25 LATE evening — Action Queue re-validation pass (Findings #232 + #233)

After §12 + §13 landed, the operator asked for a full re-validation of the Action Queue against the post-backfill universe data. The re-pull produced several material updates that this section documents (the canonical finding narratives are in docs/FINDINGS.md Finding #232 + Finding #233).

14.1 Headline shift (vs. handoff doc state)

The +$106 net P&L improvement came from 7 new trades plus the addition of TXN +$92 (Apr 22 — LONG large NULL-regime, big winner) and a few smaller contributors. The strategy is now meaningfully positive on a per-trade basis (+$1.44/trade) but still well below the +$5 to +$10/trade target.

14.2 The broken stop_triggered field — Finding #233

The re-validation immediately surfaced a critical data-hygiene bug: the trade_audit_log.stop_triggered boolean undercounts at 1.7 % (2 of 116 closed trades flagged true) while the canonical exit_reason ILIKE '%stop%' shows the real stop frequency at 49 % (57 of 116). Cross-tab:

So 55 of 57 stop-exits set stop_triggered=false. All historical analyses depending on the boolean are wrong by orders of magnitude. This invalidates several claims in the pre-Apr-25 docs:

• v2 #6 (3-stop session breaker) was claimed to have "no historical evidence" — actually 8 historical sessions with ≥3 stops (Feb 19 had 5).
• v3 D ON-HOLD reasoning ("stops eat winners on Mid-cap, RuleB will fix this") was tentative — actually strongly vindicated: Mid 80-90 % has 51.6 % stop rate, Mid 90-95 % has 86.7 % stop rate (massacre).
• v3 E ON-HOLD reasoning was based on claimed sideways-SHORT 67 % stop rate — actually 83.3 %.
• RuleB urgency was understated — corrected stop bleed is −$676 across 57 stops (avg −$11.86/trade), erasing ~80 % of the +$843 contribution from non-stops.

Forward-fix: every analysis must use exit_reason ILIKE '%stop%' until the audit-log writer is patched to set stop_triggered correctly. Canonical: docs/FINDINGS.md Finding #233.

14.3 Per-rule re-validation (4-check gauntlet on fresh data)

Rule #1 — v3 B (Skip downtrend LONGs) → KEEP rank 1, STRONGER

• ✅ n=22, all 22 distinct tickers (no de-dupe needed)
• ✅ Ex-top-2 by |pnl| = −$175 (more negative than headline because the top-2 by |pnl| includes SWBI +$70 winner masking the bucket)
• ✅ Top losers across 8+ tickers (BABA, GME, KSS, NKE, PDD, ASAN, VSCO, BBY)
• ✅ Spillover positive (uptrend +$71, sideways flat)

Verdict: PASSES STRONGER than docs claim. Lift estimate revised UP from +$140 to +$140 to +$200 / 100.

Rule #2 — RuleB (±90 min stop window) → PROMOTE rank 4 → 2

The corrected stop classification confirms the simulator's 51-stop training population matches the live audit log's 57 stops (small difference expected from reconstruction). RuleB's +$123/100 estimate is credible. Stop bleed dominates the negative contribution to strategy P&L → RuleB is the highest-impact stop-loss intervention.

Rule #3 — v2 #6 (3-stop session breaker) → KEEP IN-QUEUE rank 3

8 historical ≥3-stop sessions: Feb 19 (5 stops, −$35), Mar 2 (4 stops, −$59), and six 3-stop days. Total stops on those 8 days: 27, costing −$266. Real-world fire frequency confirmed.

v3 F (BMO + VIX sizing) → DROP to RETRACTED

Deep bias gauntlet:

The +$100 winner bucket is concentrated in STT +$37 / VZ +$33 / ALLY +$32 = +$102. Drop those 3 → 17 trades sum to −$1.58.

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 +$60 docs claim. And that's optimistic (assumes linear scaling).

Operator decision Apr 25 LATE evening: Drop both arms. Re-test the scale-DOWN arm only at trade #150+ with fresh data.

v3 D (Mid-cap LONG ≥80%) → KEEP ON-HOLD, REASONING REINFORCED

Per the corrected stop rates:

The "stops eat winners" hypothesis is now strongly supported. RuleB likely closes the gap. Note: Large 80-90% degradation suggests the future re-frame may be "Skip ALL LONGs ≥ 80%" (not just Mid).

v3 E (non-downtrend SHORT) → KEEP ON-HOLD, REASONING REINFORCED

Sideways SHORT 83.3 % stop rate (corrected from claimed 67 %) strongly supports the stops-eat-winners mechanism. KEEP set is 2-of-3 outliers (GPRO +$111, STUB +$58, GIS +$3.64). Total SHORT n=18 still below n≥30 commit threshold.

14.4 The 5-outlier rename (was 3-outlier)

Memory §K + the four-check-gauntlet docstring updated Apr 25 evening: the named outlier set extended from GPRO + DELL + SWBI (~$260 of $61 net) to GPRO + DELL + SWBI + TXN + STUB (~$430 of $167 net on 116 trades; ex-top-5 the strategy is −$263 across 111 trades).

The TXN trade also exposed a small data-hygiene gap: 5 of 116 closed trades have NULL regime_at_entry despite TXN entering on a normal trading day. Tracked in §15 (deferred — minor data-hygiene).

14.5 New observations worth tracking (not yet in queue)

• AMC + VIX ≥22 = +$128 across 8 trades — emerging sweet spot, n too small (gauntlet floor n≥12). Re-test at trade #150+.
• Mar 9 wider stops dropped frequency from 50 % → 45.5 % but raised avg loss per stop from −$9.66 → −$22.23 — the widening helped marginally but each stop is now more expensive. RuleB is the better fix than further widening.
• Large 80-90% LONG degraded from +$22 to −$3 (10 trades) — potential future widening of v3 D scope.

14.6 Final ranked queue (post Finding #232)

Combined realistic lift: +$190 to +$310 / 100 trades (slightly UP from prior +$160-$280 because v3 B's bias check came out stronger and v2 #6 is no longer purely defensive; offset by v3 F retraction).

14.7 Audit-process learnings (lesson #7 — added Apr 25 from Finding #233)

Canonical-field-validity rule. Never trust convenience flags like stop_triggered without cross-checking against the source-of-truth field like exit_reason. Audit-log boolean fields can fall out of sync silently when writer code paths drift; the textual exit_reason is the canonical record because it's set by the order-executor on every exit. When in doubt: query both, compare counts, investigate the gap.

This joins the prior lessons: 1. (§1) Use the 4-check bias gauntlet on every finding. 2. (§6 lesson #1) Sample size: n ≥ 12 minimum, ideally ≥ 20. 3. (§6 lesson #2) Skip rules need full-population spillover check. 4. (§6 lesson #3) The named outlier set (now 5) carries the strategy. 5. (§6 lesson #4) Universe view + live data must agree before shipping. 6. (§10 lesson #6 — Finding #228) True-simulator rule: proxies and observational tables are NOT counterfactual evidence. 7. (§14.7 lesson #7 — Finding #233) Canonical-field-validity rule.

14.8 Cross-references

• Canonical narrative (queue re-rank): docs/FINDINGS.md Finding #232
• Canonical narrative (broken field): docs/FINDINGS.md Finding #233
• Forward-looking source of truth: docs/STRATEGY_PLAYBOOK.md §1 + §2 + §3
• Operational mirror: dashboard 🎯 Action Queue tab (rendered from analysis.action_queue, produced by build_action_queue())
• Re-validation evidence: scripts/_revalidate_queue_evidence.sql, scripts/_revalidate_f_and_ruleb.sql, scripts/_revalidate_corrected.sql (pulled live data for this audit; safe to re-run)
• Memory rename: .cursor/memory.md §K (3-outlier → 5-outlier)
• Handoff state: docs/AGENT_HANDOFF.md §11 (current state + Action Queue)

15. Deferred from §14 (next-pass to-do)

Update Apr 26, 2026 (Phase 1 of Strategy Analysis Revamp): the first two items below have been resolved; remaining items below the divider remain deferred.

• ✅ stop_triggered writer fix — RESOLVED as Finding #236 (Apr 26, 2026). Both primary and fallback close paths in src/earnings/position_manager.py now derive stop_triggered from exit_reason ILIKE '%stop%' at write time; 55 historical rows backfilled. Boolean now matches canonical 57/57.
• ✅ 5 trades with NULL regime_at_entry — RESOLVED as Finding #234 (Apr 26, 2026). Root cause was NOT a transient SPY-fetch failure but a divergent INSERT path: TradingOrchestrator._persist_position_open was silently dropping the regime/VIX/SP500 columns while PositionManager.persist_position_open was setting them. Fixed + backfilled. regime coverage now 116/116 (100%), VIX 114/116 (98%; 2 rows Mar-18 still NULL due to yfinance ^VIX KeyError on that specific date — recoverable on next yfinance retry).

Still deferred (next-pass to-do):

• AMC + VIX ≥22 emerging sweet spot (+$128 / 8 trades) — re-test at trade #150+ with the gauntlet. Currently n=8 below the floor.
• Large 80-90 % LONG degradation (+$22 → −$3) — re-test v3 D after RuleB ships. Potential widening of scope to ALL LONGs ≥ 80 %.
• Unified simulator on universe markets (Phase 3 of Architecture Revamp) — gated by Phase 2 forward data capture (equity 1m bar collector + press- release timestamps) and the four conditions in docs/STRATEGY_ANALYSIS_ARCHITECTURE.md §0.6. (Update — ✅ Phase 3 COMPLETE: PR 1-5 shipped #306-#319, Phase 4 #320; gates met May 28, 2026.)

§15. Apr 26-27 follow-on work — DATA-QUALITY ONLY (no rule changes)

Verified Apr 27, 2026 LATE evening: the verified-findings table (§3), retracted findings (§4), 4-check gauntlet (§5), and ranked queue (docs/STRATEGY_PLAYBOOK.md §2) are all unchanged since Apr 25. Two days of sustained work between Apr 25 and Apr 27 produced zero new strategy rules and zero rule retractions. Everything that shipped was data infrastructure, audit completeness, or operational correctness.

For traceability, the per-day breakdown:

• Apr 26 — Strategy Analysis Revamp Phase 1: 6 production bug fixes (Findings #234-#239) closing data-capture gaps the v3 audit had flagged (regime/VIX/sp500 NULL, conviction_at_exit NULL, stop_triggered writer broken, dashboard ReferenceError, weekly counterfactual self-heal, long-weekend sizing reconciled). None changed strategy rules; all improved the data underlying §3 evidence.
• Apr 26 — Strategy Analysis Revamp Phase 2 + 2.1 + 2.2 + 2.x: Findings #240-#247 — 8 sub-deliverables for forward data capture (equity 1m bars hypertable + collector — load-bearing for Phase 3 unified simulator; opp_log pm_v2 snapshot enrichment; actual_beat 98.1% via post-event price-snapshot backfill; SEC EDGAR press-release timestamps 90.9%). No rule changes; the simulator that consumes them is gated on the 30-day equity-bar capture window (~May 26, 2026). (Update — that window has since CLOSED; Phase 3 PR 1-5 + Phase 4 COMPLETE #306-#322. The gating note reflects the Apr-26 state.)
• Apr 26-27 — Findings #248-#249: dashboard simplification (Phase 4 partial), Polygon mcap 2nd-tier fallback + DB-first dashboard cache cold-path. Operational quality, not strategy.
• Apr 27 — Database Integrity Audit Phases 0-6: Findings #250-#261. 6 phases of audit + remediation: silent process-failure exit-code fixes, active data-loss investigation (hold_duration_hours writer added — although that fix was itself broken until Apr 27 evening, see #263), unified integrity monitor (replaces v2/v3 watchdogs), Database tab redesign, storage hygiene + backup OOM fix, full doc refresh. No strategy changes; closes data-recording gaps that affect §3 evidence reliability going forward.
• Apr 27 evening — Dashboard 500 chain (#262): 6-commit incident resolution. Dashboard infrastructure only; no strategy.
• Apr 27 LATE evening — rigorous §B re-audit + ancillary (#263-#274): Operator requested an audit of every code commit in the last 4 days against memory.md §A-§N. Found that #256 had introduced an :exit_time::timestamptz pg8000 syntax error that silently broke the close path for every trade since deploy (hold_duration_hours, conviction_at_exit, prediction_correct, financing aggregates all NULL). Found that Yahoo's GAAP scrape had over-ridden Polymarket's confirmed YES on VZ earnings (Finding #264). Found that the dashboard volume display fix (#265) was incomplete in production. Fixed all three. Then RCA'd volume_24h_at_entry regression (Finding #272) — Polymarket's Gamma API stopped reliably returning volume24hr field; collector wrote 0; signal_generator treated 0 as None; 76 consecutive trades had NULL. Backfilled 114/117 historical rows from price_snapshots. Plus 5 newly surfaced ancillary findings (#267-#274) — most documented for operator decision rather than auto-fixed.

Net effect on the audit: none. Rule rankings unchanged. Bias gauntlet unchanged. The 5-outlier curse still defines the strategy's net P&L (GPRO + DELL + SWBI + TXN + STUB ≈ $430 of $167 net on 116 trades). Re-run scripts/strategy_performance_analysis.py for current numbers; the §3 evidence may have refreshed slightly because regime/VIX/conviction-at-exit data is now populating forward-only on every new trade.

Re-trigger conditions for next audit pass: (a) 50 more closed trades (currently 117; trigger at 150+), or (b) Action Queue items #1 + #2 (skip downtrend LONGs + RuleB ±90min) ship and we want to measure impact. (Update — (b) has shipped: Rule #1 is active (#304) and RuleB is in Phase 1 dry-run (#305); live impact measurement is signal-density-limited, which is why the Phase 3 simulator now carries the counterfactual evaluation.)

§16. Jul 5, 2026 — RuleB retraction (Finding #330): the simulator's window arm was gap-fill optimistic

Trigger: operator-directed full strategy review (Jul 5, 2026). Three independent observations converged:

1. All four June 2026 in-window stops (LULU, ADBE, ACN, KR) were CORRECT stops — the price kept falling through the window. Holding them to the rearm/scheduled exit would have been worse in every case (ACN: −$26.10 actual stop vs ~−$134 at the 10:32 scheduled exit through the post-earnings crash).
2. A gap-aware ad-hoc replay of the RuleB arm flipped its sign.
3. Tracing simulate_window_only confirmed the bug in code: after the silent window, any bar with low <= stop returned an exit at the stop price, even when the price had crossed the stop inside the window and was still beyond it at rearm. The as-shipped #305 cancel-and-rearm mechanism re-places a STP-LMT that triggers instantly and fills near market — simulate_finding_a_hard already modeled this as rearm_gap_open (the arm that retracted FA_hard at −$1,545/100 in §10); the RuleB arm never got the same handling.

Fix + corrected numbers (prod re-run, n=129): simulate_window_only is now gap-aware (first post-window bar OPEN fill when the open is already beyond the stop; signature unchanged so the phase3_simulator reference-adapter contract holds; RuleG inherits).

Optimism bias ≈ $535 across 129 trades. Ex-repeat-tickers RuleB is −$226.67/100 — negative on every §K check. Mechanism: the window's recovery saves (dip-through-then-recover, preserved in the gap-aware sim) are outweighed ~2:1 by gap-through cases where the rearm fill is far worse than the original stop fill.

Decisions: RuleB Phase-2 active flip CANCELLED (Phase-1 dry-run code remains, logging only — its 8,026-duplicate-line log spam also fixed); v3 D + v3 E RuleB dependencies REMOVED (§12.2 / §12.3 re-evaluate as pure selection rules at trade #150+ / SHORT n≥30); the #323.b "0.92-0.99 conditional on RuleB active" lever is dead. NO stop-loss rule variant currently survives the gauntlet.

Audit-process lesson #7: a TRUE simulator (memory §K) must model the REARM path of any cancel-and-rearm rule, not just the silence. The PR-1 §10.1 cross-check (simulator-vs-reference consistency) cannot catch a shared mechanism error — mechanism review against as-shipped order semantics is a separate mandatory step for any stop-manipulation rule.

Canonical narrative: docs/FINDINGS.md Finding #330.