Monitor

Execution-realism integrity audit (see completedItems, same date) found that under either execution-realistic entry variant (first-actionable-candle close, or canonical price only if later revisited), the ALREADY-APPROVED 'quality' tier component of the frozen V2 architecture (candidate-architecture-v2-tiered-quality-and-veto, Component 2) loses its positive edge entirely (+0.2508R canonical -> -0.0461R / -0.0289R realistic, a sign flip, n=113). The 'elite' tier (Component 3) and the PDH-veto (Component 4) both survive directionally (shrunk 33-80% but still correctly signed). The pooled/aggregate population also sign-flips (+0.1788R -> -0.0672R / -0.0246R, n=495).

Per AJ's instruction to freeze all promotion decisions and rebuild the validated architecture under Variant B (primary)/Variant C (secondary) causal execution, re-ran the full leak-free walk-forward ledger (same population construction as candidate-architecture-v2, WARMUP=50) plus a full hardening battery on each population. RESULT (classification B -- only a narrow subset survives): pooled/baseline population collapses to statistical noise under Variant B (exp +0.1952R->+0.0039R, PDH-PASS one-sample CI crosses zero) and is outright negative under Variant C. The 'quality' displacement-only tier is weak and not robust under either variant (~81% degraded under B, not significant and sign-flipping on recency under C). The displacement+PDL-swept 'elite' population remains genuinely robust under Variant B (exp +0.4173R, permutation p=0.0016, block-bootstrap CI excludes zero, survives LOYO/top-5-removal/warmup sensitivity/recency split) but is materially more fragile under the stricter Variant C (p=0.0851, not significant; 81.1% of total R concentrated in its top 5 trades; 2026 turns negative) -- so even this surviving subset should be read as 'narrow and execution-sensitive,' not as a settled edge.

Per AJ's item 10 (this same turn): open a future implementation item so a production READY alert can never advertise a stale confirmation-bar entry price, given the execution-realism audit confirmed both (a) the confirmation-bar price is frequently not the earliest-actionable price (lag >0 in 60.2% of Phase C trades) and (b) lib/live/notifications.ts's READY push body reads `setup.confirmationClose.close` directly with no current-market-price check at send time. Explicitly NOT implemented this turn -- AJ's instruction was to queue this item only.

Independent Strategy Family #1 (Liquidity Sweep -> Failed Breakout / Reclaim) is complete -- pre-registered manifest at research-output/liquidity-sweep-reclaim/CAMPAIGN_MANIFEST.md (commit 6348e9f) + PHASE0_DATA_CAUSALITY_AUDIT.md, full synthesis in research-output/liquidity-sweep-reclaim/FINAL_REPORT.md, raw results in phase3to7_results.txt. This is a wholly new, non-V1 setup family (level sweep -> failed breakout -> close-based reclaim -> next-bar-open entry), causally reconstructed from MarketCandle history with no V1 condition/displacement/inversion vocabulary reused. One mechanical, pre-outcome bug was found and fixed during execution (funnel-counter over-counting for consumed levels, ~300x inflation, confirmed to NOT affect the actual traded population -- commit a95bdcd), logged per manifest Sec. 10, not a definitional change. Full-history scan (2,518 executed trades, 2022-06-03..2026-09-10, ~503.6 trades/yr -- frequency was never the limiting factor): Phase 3 BASE (fixed 2R target, no filters) is dead flat (expectancy=-0.0016R, PF=0.998, block-bootstrap 95% CI [-0.0634, 0.0604] straddles zero, sign-test p=0.966). Phase 4: ZERO of 10 pre-registered mechanism dimensions survive Bonferroni correction (family alpha=0.005); none even reach nominal significance except two borderline continuous measures (time/closes beyond level before reclaim, p=0.06/0.04) that still fail correction. Phase 5: both structural exit variants (opposing-range, nearest-opposing-liquidity) are substantially WORSE than the flat baseline (exp=-0.145R and -0.055R respectively, driven by 5.8-6.2% win rates). Phase 6: not eligible (zero Phase 4 survivors, so no combination architecture was even tested). Phase 7: chronological half-split and leave-one-year-out both show the sign flipping depending on which slice is examined (exactly the signature of a null population, not a decaying real edge); top-winner removal barely moves the number; a full parameter-neighborhood sensitivity sweep (4H-swing lookback 10/20 vs 15, acceptance timeframe 15M vs 5M, observation horizon 180/300min vs 240min -- 5 variant re-scans, all via CI dispatch) reproduces the same near-zero result in every variant (expectancy range -0.0027R to +0.0046R, PF range 0.996-1.007, every bootstrap CI straddling zero, frequency actually rising to 560-741/yr under looser variants with no added edge) -- the null result is robust to the exact parameter choice, not a fragile artifact of the specific pre-registered defaults. FINAL CLASSIFICATION: D -- FAMILY FAILS. Manifest Sec. 9 prospective/shadow logging (PB-context feature values on this new population, firewalled from every Phase 3-8 statistic) was executed as part of the Phase 2 scan itself (all 2,518 events carry pbShadow metadata, 87-100% field coverage) -- interpreted as satisfying AJ's 'parallel research-only prospective logging' instruction within the manifest's own pre-registered scope (Sec. 0: lives entirely under lib/strategy-engine/research/sweepReclaim/ and scripts/, read-only against MarketCandle); no live-pipeline integration was built, since the manifest's own standing scope statement commits this campaign to historical read-only analysis and a D classification creates no candidate finding that would need forward validation -- this scope choice is stated explicitly for AJ to override if genuine forward-in-time logging is wanted regardless of outcome. No V1/V2/production code touched; no sizing performed; no research finding promoted to live execution; no merge to main of research/strategy code. Two small, transparently-flagged CI workflow-definition changes were merged to main (PR #80: add the scan workflow; PR #81: plumb sensitivity-sweep params through it) -- read-only CI plumbing required to dispatch this campaign's own scans, not strategy logic.

Independent Strategy Family #2 (Breakout Acceptance -> Retest/Hold -> Continuation) is complete -- pre-registered manifest at research-output/breakout-acceptance-continuation/CAMPAIGN_MANIFEST.md (commit 4630541) + PHASE0_DATA_CAUSALITY_AUDIT.md, full synthesis in research-output/breakout-acceptance-continuation/FINAL_REPORT.md. Deliberately the mirror-image mechanism of Family #1 (break -> accept -> retest -> hold -> continue, vs. Family #1's break -> fail -> reclaim -> reverse); reused Family #1's already-proven LevelTracker and PB-shadow firewall verbatim (generic causal level arithmetic, not either family's own strategy logic), built an entirely new state machine (BreakoutContinuationEngine) with 3 pre-registered acceptance-definition variants (A0 single 5M close primary, A1 two consecutive closes, A2 ATR-normalized min distance) and 3 entry arms (A=acceptance, B=retest/hold, C=no-retest late continuation). Per AJ's explicit efficiency directive, all 3 variants were extracted in ONE causal forward pass each within a SINGLE CI job/DB fetch (~5 min total for the full 2022-06-03..2026-09-14 history) -- deliberately avoided Family #1's own concurrent-push race by never parallel-dispatching scan jobs. Cross-validated against Family #1's identical shared primitives (touchedAndBroken/duplicateLevel counts match almost exactly; acceptedCount+immediatelyRejected+breakoutExpiredUnresolved = touchedAndBroken-duplicateLevel exactly, both in the 6-week validation window and full history) before trusting any downstream statistic. RESULTS (n=1,353 ACCEPTED events, 1,345/975/331 executed trades in Arms A/B/C respectively): Arm A (acceptance-only entry, 269/yr) is directionally positive at the primary FIXED_2R metric (+0.0402R) but NOT robust -- bootstrap CI crosses zero, LOYO flips sign excluding 2025, half-year expectancy oscillates both directions with no trend. Arm B (retest/hold required, 195/yr) actively HURTS -- significantly NEGATIVE at FIXED_1R (p=0.0027), directly answering AJ's Q2 that requiring a retest hurts rather than helps. Arm C (late entry only after confirmed NO retest within horizon, 61-66/yr) is the one credible specialist lead: FIXED_3R is robust across every check (LOYO all-5-years-same-sign, PF 1.31-1.49 in every year-exclusion, top-10-winner-removal still strongly positive, half-year positive in 8/10 buckets), and FIXED_2R (the pre-registered primary) is borderline at the exact 240-minute default (CI=[-0.0249,0.3411], barely crosses zero) but reaches full significance under a modestly widened (not narrowed) 300-minute retest-horizon sensitivity check (both bootstrap CIs excluding zero), with the effect strengthening rather than weakening under that neighborhood perturbation. Phase 6: 3 of 11 mechanism dimensions survive Bonferroni correction (minutes-to-first-retest, failed-retest-attempt-count, pre-entry continuation-strength -- all describing retest-hold conviction). Phase 7: zero of 7 context dimensions survive; one (intraday-trend-alignment) is mechanically degenerate by construction (a level broken after the session open structurally requires price already on that side of the open -- verified directly against raw event data, confirmed not a bug). Phase 8: both structural exits (nearest-opposing-liquidity, measured-move) underperform the flat fixed-R baseline. Phase 9: the one economically-coherent combination of the two related Phase 6 survivors (zero failed attempts AND top-tertile continuation strength) COLLAPSES rather than compounds (exp=-0.1061R vs parent Arm A's +0.0402R) -- a real finding, not a rescue. FINAL CLASSIFICATION: C -- SPECIALIST EDGE ONLY (Arm C is the specialist edge at 61-66 trades/yr; Arms A and B do not clear the bar; no broad BASE exists; no combination rescues anything). PB-shadow prospective logging executed for all 1,353 events (87-100% field coverage), firewalled from every Phase 5-12 statistic per the same convention as Family #1. No V1/V2/production code touched; no sizing performed; no research finding promoted to live execution; no merge to main of research/strategy code (one CI-workflow-definition-only PR, #82, merged to main -- read-only plumbing, not strategy logic).