Research History

PHASE C COMPLETE: all 28 chronological backfill chunks (2023-07-19..2026-09-10) done and verified directly against production. Final totals: 379 ResearchCandidate, 5759 ResearchCandidateFull, 3450 RegimeSnapshot rows. Every chunk has non-zero rows in all three tables; every chunk with an individually-confirmed count matches exactly. Nine infra blockers found and fixed along the way (BLOCKER-1 through BLOCKER-9); see… — Artifact: PR #77 (verification workflow, merged to main); scripts/verifyPhaseCBackfillComplete.ts; run 34734671624

research-output/campaign-state.json

FORENSIC AUDIT (per AJ's explicit pause instruction): investigated AJ's observation, via the new visual replay UI, that many MNQ 2026-08-27 bullish candidates reaching Sweep->Displacement->Inversion appeared to terminate for 'Valid target already hit before entry.' CONCLUSION: A -- correct current V1 behavior, no engine bug. Ground truth (Part A, canonical Phase C chunk28 data, engineVersion 3b6e746): all 14 bullish… — Artifact: scripts/auditTargetAlreadyHitAug27_partA.ts (commit 2099c7f, run 34763134232); scripts/auditTargetAlreadyHitAug27_partB.ts (commit 0693e74, run 34763174915); scripts/auditTargetAlreadyHitAug27_partC.ts (commit a9abacb, run 34763316490)

research-output/campaign-state.json

AUDIT CORRECTION: during release-preview validation of the above forensic audit, AJ flagged an apparent parity issue -- the release preview showed the canonical READY candidate (setup_MNQ_1787835600_5001) decision timestamp as 9:42 AM ET, while the forensic audit's own chat table had reported that candidate's touch/sweep/confirm as 09:20/09:50/10:22 ET. Investigated with a programmatic (zoneinfo-based, non-manual) r… — Artifact: research-output/AUG27_FORENSIC_TABLE_CORRECTED.md; research-output/build_forensic_table_v2.py

research-output/campaign-state.json

EXECUTION-REALISM INTEGRITY AUDIT (AJ, escalated to a separate priority audit after the timestamp-semantics audit closed as A/no-lookahead): audited the full official V1 Phase C population (495 executed/READY trades, all years) for whether the recorded confirmation-bar entry price/time is systematically earlier than the candle on which the full gate sequence (sweep+displacement+inversion+trap-clear, i.e. the CHECKIN… — Artifact: scripts/auditEntryTimingRealism.ts (commit 1307168, run 34768159096)

research-output/campaign-state.json

CAUSAL-EXECUTION ARCHITECTURE REVALIDATION (per AJ's explicit instruction to freeze all promotion decisions, treat canonical V1 historical results as reference-only, and rebuild the validated architecture under Variant B [first-actionable-1M-close entry, primary] with Variant C [revisit-gated, secondary] as the execution-realistic benchmarks going forward). Replicated the EXACT walk-forward population construction (… — Artifact: scripts/auditCausalArchitectureRevalidation.ts (commit 46f48ef, run 34768759090, job 103754376828)

research-output/campaign-state.json

EXECUTION-REALISTIC STRATEGY DISCOVERY CAMPAIGN, PHASE 0-2 (AJ's new 10-phase mandate; this turn covers Phase 0 dataset integrity, Phase 1 from-scratch hypothesis revalidation, Phase 2 failure diagnosis only, per AJ's explicit 'do not proceed to Phase 3 until I approve' instruction). PHASE 0: audited the FULL Phase C funnel (5759 rows, not just the 495 executed trades) -- terminal states NO_TRADE 51.2%/EXPIRED 29.8%… — Artifact: scripts/execCampaign2_phase0_datasetIntegrity.ts (commit c5cbf58, run 34769622558, job 103756709398); scripts/execCampaign2_phase1_hypothesisRevalidation.ts (run 34769626559, job 103756720543); scripts/execCampaign2_phase2_failureDiagnosis.ts (run 34769631804, job 103756736464)

research-output/campaign-state.json

BASE/FRONTIER SEARCH -- PHASE 1b (AJ's explicit follow-up: the 46-trade Elite population alone does not satisfy the research objective; investigate whether a broader causal-execution positive-expectancy BASE exists, produce an edge-vs-frequency frontier rather than picking the single highest-expectancy tiny subgroup, report annualized opportunity counts, and do not manufacture BASE by relaxing filters). Screened all… — Artifact: scripts/execCampaign2_phase1b_baseFrontierSearch.ts (commit a424d9b, run 34770139115, job 103758115504)

research-output/campaign-state.json

FINAL CURRENT-FAMILY RESEARCH BEFORE NEW STRATEGY FAMILIES (AJ's explicit follow-up: harden the pdlSweptAtDecision-alone lead, resolve execution economics with predefined causal models, and define PB-style concepts for future review -- before permanently classifying the current setup family as Elite-only). ITEM 1 (PDL-alone hardening, unchanged definition, full Phase-1 battery): under VARIANT B (primary): n=95, exp=… — Artifact: scripts/execCampaign2_phase1c_pdlAloneHardening.ts (commit 58257e0, run 34770917946, job 103760235974); scripts/execCampaign2_phase3prep_executionModels.ts (run 34770922089, job 103760247065)

research-output/campaign-state.json

PB CONTEXT RESEARCH -- HYPOTHESIS 1: structural manipulation raid vs. condition-owned local sweep (AJ's single, pre-registered hypothesis, per the PB concept-definition review's item 3). Pre-registered 'sweep extent' (extreme price from sweepReferenceWickTime through sweepConfirmedAt) and 4 raid definitions BEFORE running: Variant A = PDH/PDL raid via flip-detection (calls the existing, unmodified computeRegimeSnaps… — Artifact: scripts/execCampaign2_pbHyp1_structuralManipulation.ts (commit 68573a7, run 34772113655, job 103763477181)

research-output/campaign-state.json

PB CONTEXT RESEARCH -- HYPOTHESIS 2: Higher-Timeframe Draw on Liquidity (HTF DOL) (AJ's second, and per this turn's explicit stop condition, final PB hypothesis for now). PHASE 0 AUDIT (per AJ's explicit 'do not assume the legacy field is suitable' instruction): inspected target.ts's evaluateTargets()/extendedDOL infrastructure directly. Findings: (1) recognized levels -- tiers 1-3 are condition-timeframe + 15M/1H p… — Artifact: scripts/execCampaign2_pbHyp2_htfDrawOnLiquidity.ts (commits 22d1c66, fe9adf6; cancelled run 34772997072; successful run 34775962497, job 103773995886)

research-output/campaign-state.json

PB CONTEXT RESEARCH -- PHASE 2B: Daily/4H Swing-DOL Hardening (AJ's follow-up: harden and diagnose the ONE frozen swing-DOL hypothesis carried forward from Phase 2 -- do not reopen the failed general-DOL/Daily-FVG/4H-FVG types, do not combine with PDL/Elite/displacement/other filters, do not optimize a distance threshold). Pre-registered 6 confirmatory comparisons (overall Aligned-vs-Not, Daily-swing-marginal, 4H-sw… — Artifact: scripts/execCampaign2_pbHyp2b_swingDolHardening.ts (commit ccaef92, run 34777283264, job 103777600755)

research-output/campaign-state.json

PB COMPREHENSIVE CONTEXT STRATEGY RESEARCH CAMPAIGN (AJ's explicit directive: one comprehensive, pre-registered, controlled campaign covering all remaining PB-style concepts in one pass, rather than one-hypothesis-at-a-time with a stop after each -- Phases 0-13 per AJ's own numbering). PHASE 0: committed research-output/PB_COMPREHENSIVE_CAMPAIGN_MANIFEST.md BEFORE any result was read -- exact definitions, population… — Artifact: research-output/PB_COMPREHENSIVE_CAMPAIGN_MANIFEST.md; scripts/execCampaignPB_phase1to4.ts (commit 9849671, perf-fix ca0d819, run 34781920664, job 103790374773); scripts/execCampaignPB_phase5to9.ts (commit 50e9c09, perf-fix ca0d819, run 34782993090, job 103793279311); scripts/execCampaignPB_phase11combination.ts (commit 84f5f7f, run 34783544478, job 103794787586); research-output/pb-comprehensive-campaign/FINAL_REPORT.md; PR #79 (timeout increase, merged to main)

research-output/campaign-state.json

Phase B complete: proved computeRegimeSnapshot/persistRegimeSnapshot against real historical data (not just synthetic fixtures). Replayed 2022-06-03..2022-07-14 (chunk1) via the same bounded-bootstrap/StrategyEngine mechanism the official audit routes use, computed+persisted a RegimeSnapshot for every setup with a confirmed entry. 76 of 147 total setups reached a confirmed entry -> 76 rows persisted, DB count verifi… — Artifact: scripts/proveRegimeSnapshotSmallSample.ts; .github/workflows/prove-regime-snapshot-small-sample.yml; run https://github.com/meatham07-design/TradeRep/actions/runs/34544549141

research-output/campaign-state.json

Made and recorded the Phase C scoping decision, unblocking the major historical backfill: chronological 6-week chunks over the untouched 2023-07-19..2026-09-10 window (28 chunks total). Extended the pre-existing, never-yet-wired generic backfill runner (scripts/runResearchBackfill.ts) to persist ResearchCandidate + ResearchCandidateFull + RegimeSnapshot from a single replay per chunk, reusing already-computed struct… — Artifact: scripts/runResearchBackfill.ts; lib/strategy-engine/research/dataset/persistFullFunnelCandidates.ts (added rows to PersistFullFunnelResult); .github/workflows/run-research-backfill-chunk.yml

research-output/campaign-state.json

Root-caused and fixed the real warmup performance bug behind Phase C chunk1's failure (BLOCKER-2): detectPivots() re-scanned the entire closed-bar history from index 0 on every one of ~8 calls/setup, the one piece of the liquidity-pool pipeline that stayed uncached even after aggregateClosedBars was already made incremental in an earlier fix for the identical symptom (see that function's own doc comment on the OOS21… — Artifact: lib/strategy-engine/research/liquidity/pools.ts; tests/unit/strategy-engine/research/liquidityPoolsIncrementalPivots.test.ts; commit 5e5d629

research-output/campaign-state.json

IMPORTANT CORRECTION to the entry above: chunk1's retry with the detectPivotsIncremental fix still crashed after 1h55m -- a genuine JS heap OOM at ~6.36GB, dying at essentially the same historical point (2023-04-24) as the original failure (2023-04-20), with the same climbing/noisy per-session timing. The pivot-detection fix was real and independently verified, but it was NOT the actual root cause of the crash. Conf… — Artifact: run https://github.com/meatham07-design/TradeRep/actions/runs/34564811529 (job log showing FATAL ERROR: JavaScript heap out of memory)

research-output/campaign-state.json

Fixed BLOCKER-3 (the real cause of chunk1's OOM): added an opt-in pruneOutOfRangeResearch flag to computeHistoricalAuditReport that nulls a terminal setup's .research once its session falls behind fromDate. Safe because nothing ever reads .research for a setup outside inRangeSetups -- traced every consumer (persistFullFunnelCandidates, buildLiquidityAuditReport, buildRuleCompliantLiquidityReport, buildRuleCompliantN… — Artifact: lib/backtest/audit/computeHistoricalAuditReport.ts; scripts/runResearchBackfill.ts; tests/unit/backtest/pruneOutOfRangeResearch.test.ts; commit 202bf55

research-output/campaign-state.json

Chunk1's 4th attempt (all four then-known fixes applied) SUCCEEDED with real data (14 ResearchCandidate + 248 ResearchCandidateFull + 143 RegimeSnapshot rows). RegimeSnapshot fix confirmed in production (225.6s actual vs 228s benchmarked -- exact match). CORRECTION: the displacement-ATR bound fix did not meaningfully help in production (12,127.0s replay vs the pre-fix 11,502.6s -- essentially unchanged, same climbin… — Artifact: lib/strategy-engine/aggregation/timeAggregate.ts; tests/unit/strategy-engine/timeAggregate.test.ts; run https://github.com/meatham07-design/TradeRep/actions/runs/34602950737 (chunk1 success log)

research-output/campaign-state.json

Corrected V1 nearest-untapped-liquidity TP rule implemented, verified (tsc/eslint/full suite 748/752 passing, 4 pre-existing unrelated webhook-test failures), merged into claude/research-first-cross-entry — Artifact: PR #71, commit 34fe74b

research-output/campaign-state.json

Closed PR #70 (chatgpt/v1-nearest-untapped-tp) as superseded/unverified in favor of #71, after persisting a concrete equivalence/verification-gap proof — Artifact: PR #70 (closed); docs/pr70-vs-pr71-equivalence-proof.md

research-output/campaign-state.json

Unified the previously-separate ML-quality roadmap and the new regime-robustness directive into one locked reference doc, reframed as one pipeline (parallel setup-quality/regime tracks, not sequential bots) with ongoing post-deployment monitoring explicit — Artifact: docs/TRADE_RESEARCH_ROADMAP.md

research-output/campaign-state.json

Created campaign working branch off the corrected baseline — Artifact: claude/regime-robustness-campaign (base: claude/research-first-cross-entry @ 34fe74b)

research-output/campaign-state.json

Built and proved end-to-end the non-blocking campaign orchestration controller (queue-driven, dispatches next mechanical task via Actions API). Discovered and solved the real platform constraint: workflow_dispatch/schedule only work for files registered on the default branch, so the controller + verification workflows (infra-only, zero strategy code) were merged to main via #72. First live test: manually fired the c… — Artifact: PR #72 (merged); .github/workflows/campaign-controller.yml; scripts/campaignController.ts; docs/TRADE_RESEARCH_ROADMAP.md Part 10

research-output/campaign-state.json

Established an hourly self check-in (Routine, self-bound to this session) to poke the controller and service analysis-kind queue entries without waiting for AJ to prompt. Confirmed via one manual fire that it retains full tool access (an earlier 'no connectors' warning at creation time did not apply to this persistent-session self-bind mode). — Artifact: trigger_015QVdJgXAkJS8jJEfFFfnAp, hourly at :53

research-output/campaign-state.json

Added a read-only MarketCandle coverage report to resolve BLOCKER-1 with evidence rather than assumption, after noticing a prior 'all-years' historical import workflow had already succeeded on 2026-09-06 (far more than the ~13 months of REPLAY/research coverage this ledger previously conflated with import coverage) — Artifact: PR #73 (merged); scripts/reportMarketCandleCoverage.ts

research-output/campaign-state.json

CONFIRMED the corrected V1 baseline actually produces trades (AJ's original complaint). 2022-06-03..2022-08-25 (52 sessions): 332 setups, 41 accepted trades, 21W/20L, total R +3.02, total $ +111.50, and 'target already hit before entry' rejections are 0 -- the old bug is gone. Took 3 iterations of the verification workflow itself to get a readable result (OOM masked by a piped exit code, then --json output too dense… — Artifact: research run https://github.com/meatham07-design/TradeRep/actions/runs/34528678859

research-output/campaign-state.json

Phase B schema design: audited existing schema first and found ResearchCandidateFull already satisfies the Setup Feature Dataset (Part 4) verbatim -- reused as-is, not rebuilt. Added RegimeSnapshot, a new additive table for the actual gap (Part 5/6: environment/regime + ATR-normalized measures), 1:1-paired with ResearchCandidateFull via (sourceRunLabel, setupId). Migration is a single CREATE TABLE, no ALTER on any e… — Artifact: prisma/migrations/20260910215622_add_regime_snapshot/; docs/phase-b-schema-design.md; tests/unit/strategy-engine/research/regimeSnapshotSchema.test.ts

research-output/campaign-state.json

Built computeRegimeSnapshot (pure, causal-only extraction covering every RegimeSnapshot field group: volatility, trend/chop, overnight/session structure, time context, ATR normalization) + persistRegimeSnapshot (writer). Reused existing proven pieces (atr() from displacement/metrics.ts, the CME 18:00-ET rollover convention already used for daily bias) rather than reinventing them. Writing tests caught two real bugs… — Artifact: lib/strategy-engine/research/regime/computeRegimeSnapshot.ts; lib/strategy-engine/research/regime/persistRegimeSnapshot.ts; tests/unit/strategy-engine/research/regimeSnapshotExtractor.test.ts

research-output/campaign-state.json