🐛 Incidents + 🚀 Enhancements per Asset Class

git grep finds the live DB_PASS_STOCKS value stocks1234560 committed in 5+ files on main + reproduce-blocks of open PRs #481/#482/#485. /updates docs publish to the public site. Pre-existing widespread leak.

compute_pf() at line 47-50 returned wins/losses (count ratio) with 999 fallback, not sum(pnl_wins)/abs(sum(pnl_losses)). All claimed PF 600+/400+/200+ from mutation scan were math artifacts. Fixed to gross_profit/abs(gross_loss).

compute_pf returned wins_count/losses_count with 999 fallback when losses==0, not gross_profit/gross_loss. Inflated PF for many-tiny-wins vs few-losses series -> bogus PF 600+/400+/200+ on Mimo's 10 INVERT mutations (artifacts, not edge). Flagged P0 by peer gx10, verified on disk by claude-opus-4-8.

Mutation framework compute_pf is broken: returns win/loss COUNT ratio, not Profit Factor

DSR_PARAMS["n_trials"] was None, _load_real_n_trials() never called. TypeError on n_trials <= 1 comparison.

rigorous_backtest_harness.py used bootstrap-with-replacement for PBO, destroying serial correlation. Momentum/trend strategies always appear insignificant. Fixed with moving-block bootstrap.

Orphan test file blocking all PR CI: test_money_ready_capital_lock referenced symbols not on main

money_ready_verdict 2026-06-02 10:19Z: every class NOT_READY/INSUFFICIENT. CRYPTO PF0.92/WR36%, EQUITY PF0.33/WR27%, FOREX PF0.48; ETF/COMMODITY 3-4 trades only; BOND n=0. Research/paper edge (ai-tournament deepseek_v4 PF3.46) does not survive into policy-clean layer.

21 picks across alpha_engine/data/closed_picks.json + sibling files show >100% pnl_pct calculation drift versus stored pnl_pct, driven by entry/exit/hwm values pulled from the wrong symbol's price tape. SHIBUSDT: entry=5.53e-06 (correct), exit=4100.97 (off by 9 orders of magnitude — likely SHIB-stoc…

open_bloat check on 2026-05-25: db_health_check.py queried bt_backtest_trades (millions of backtest rows) and reported 29,254,204 OPEN rows. The incident was incorrectly attributed to trading_picks (which had ~46K rows at the time). info_schema estimate for bt_backtest_trades was 1,271,867 — the 23x…

The current audit pipeline shows rejects in aggregate but provides no durable lane for picks that failed gates and later would have won materially. That hides false negatives and prevents learning whether concentration, thin-sample, or quarantine rules are discarding real edge.

The High Conviction decision path is split across audit_dashboard/hc_filter.js and tools/dashboard_hc_rules.py. EAGLE review found likely drift around confidence handling and small-sample relaxations, so the same pick can qualify differently depending on which surface evaluates it.

Ring-2.6-1T noted these numbers come from cell-by-cell search across (confidence × R:R × strategy family) buckets — they are not actionable forward signals but the /audit SUPREME EDGE callout presents them as if they were. 3/3 swarm engines (deepseek + cerebras + gemini) confirmed REAL P0. Risk: a u…

Distinct from the known confidence-anti-predictive incident: the ranker formula itself bakes the inverted signal into the headline score with the largest weight. Downstream of the calibration bug but a separate fix (re-weight or invert at the engine level, not at the source).

Opencode 2026-05-12 identified the missing upstream writer that should read ACTIVE at_raw_picks, detect TP/SL/time-exit per asset class, and feed new entries into closed_picks.json. Without it the signal_outcomes table has 0.09% coverage of raw picks — every forward-WR claim is built on 0.1% of the …

won_pnl_contradiction check: 2,531 rows tagged status='WON' have avg_pnl=-41.13%, 9 with negative pnl. SL_HIT rows are all negative as expected (good); TP_HIT all positive (good); LOST rows mostly negative (correct). The WON status is a labeling bug, not a stats artifact — every claim using status='…

open_bloat check: 29,254,204 open status rows. info_schema estimates 1,271,867 — actual count 23x the estimate. Last terminal write was 2026-05-12 23:42 (270 hours ago). The forward_validator is frozen — no picks have been closed in 11+ days.

cot_paper_pilot.py counts the same weekly CFTC release as ~100 separate trades. Inflates n from ~5 real unique releases to 101. The DSR=1.0/WR=86.5% headline is therefore overstated. Three independent AI audits flagged this.

data/smart_picks.json last regenerated 2026-04-30T02:56. The dashboard reads smart_picks_feed which IS more recent (~1.5h), but the underlying picks may be cycled with stale entry prices.

Unit-clamp bug commit #876 missed 5 rows. Distorts FOREX avg to -8% and rounds PF to 0.00, making the entire class look catastrophic even though baseline WR is 43.9% on n=1666.

Confidence is anti-predictive: conf>=0.9 -> WR 14.4%, conf 0.5-0.6 -> WR 60.3%. The 5-factor Smart Picks engine weights quality/elite_score at 35% which is derived from confidence, so the top-of-funnel ranker is structurally flipped — at least for crypto.

Last resolved 2026-03-04. Outcome resolver pipeline appears dead. All forward-WR performance claims unverifiable because signal_outcomes has only 0.09% coverage of raw picks.

176 non-crypto 'closed' rows have entry==exit/pnl=0 with exit_reason FORCE_CLOSED_TOXIC (113) or RESOLVE_FAILED_MAX_RETRIES (60), mostly multi_asset_copytrader+forex_copy_trader. FOREX: 94/138 toxic. These are price-fetch failures, NOT resolved trades; they crushed WR denominators (raw FOREX ~4% vs …

Cross-sectional attribution probe (tools/attribution_probe.py): deepseek_v4 alpha t=1.74<2.0, crowd_beta 0.49; all large-n models crowd_beta 0.7-1.0; only tiny-n flukes pass. No model has adequate-n + surviving alpha. Edge is crowd/beta not skill.

Category taxonomy case-mess: 23 'stock' + 590 'stocks' should be EQUITY; 4 'penny' + 4 'pennystock' should be EQUITY; ~394 UNKNOWN/NULL rows auto-taggable by symbol pattern

53.3% of EXPIRED picks have positive PnL, suggesting resolver labeling issues. This inflates WR and makes strategies look better than they are.

92 out of 127 systems have never produced a single resolved trade. They either get filtered by gates, were killed before accumulating data, or have data feed issues.

98.9% of generated picks are filtered by quality gates. 203 strategies permanently killed. ml_crypto_predictor generates 492 picks with 0 surviving. This starves the system of trades.

AI tournament data (5,492 picks, 3,692 resolved, top 5 T1 models) shows CRYPTO LONG has 33% WR / -0.49% avg PnL while SHORT has 67% WR / +3.74% avg PnL. Production scanner emits LONG when SHORT has edge.

Stale OPEN picks resolved: ~100k+ -> 3,806 (97% reduction) via controlled batch resolve

Sign-flips, COT look-ahead, extreme source concentration (HHI > 0.25). Classes frozen from production promotion.

Tournament data: SHORT 67% WR / +3.74% avg PnL vs LONG 33% WR / -0.49% avg. Production scanner emitting both directions.

After fixing harness: CRYPTO 46 strats (1 T3), FOREX 11 (0 T3), COMMODITY 4 (0 T3), EQUITY 1 (0 T3), ETF 0. PBO is binding constraint.

Stop-loss tightness: 626 SL hits at 0.5% WR producing -2257% cumulative PnL

ML confidence is anti-predictive: conf>=0.90 yields 14.4% WR vs conf 0.50-0.60 yields 60.3% WR

EXPIRED mislabel: 53.3% of EXPIRED picks have positive PnL (resolver mislabeling WINs as EXPIRED)

Source-system concentration: kimi_riseoftheclaw = 43.6% of all picks (HHI 0.232 > 0.20 cap)

Swarm review (reports/best_picks_swarm_review_2026-06-02.json): production aggregate NOT_READY. Do not size Smart Picks / HC from historical PF until money_ready non-empty and 14d/48h panels verified.

Resolver hygiene: OPEN count was ~100k+ before 2026-06-02 controlled batch resolve (resolve_stale_open_picks.py --execute --max-batches 5 --batch-size 500 → ~3,806 OPEN remain). Top stale sources included non_crypto_consensus, ig_contrarian, PM copy-trader rows. Continue sliced batches until OPEN al…

tools/check_resolver_health.py reports forward_test table/columns absent on live ejaguiar1_stocks. Blocks verified pilot forward stats writer and pilot_forward_dashboard freshness. See alpha_engine/mysql_trading_sync.py for ALTER SQL.

money_ready_verdict.json 2026-06-02: all classes NOT_READY or INSUFFICIENT_DATA. Policy-clean CRYPTO PF 0.92 / WR 36%; EQUITY PF 0.33 / n=52. Tournament paper edge does not flow into asset_class_health. No sizing up until forward n≥30–50 + DSR/SPA per class.

Raw DB CRYPTO 90d ~39% WR / PF 0.37 vs marketed funnel cell. Leakage: duplicate signal_ts groups, EXPIRED→WON mislabels, claude_gainer_st concentration. Apply mutate-before-kill; verify 14d/48h panels before any promotion.

~73 class x strategy cells tested with no Bonferroni/FDR. 2 surviving PF>1 at alpha=0.05 is chance-level. At alpha=0.05 over 73 cells need p<0.00068 — neither crypto sleeve clears at n~30.

Published JSON gen 2026-05-30T23:26Z has 1,037 rows / 42 models; live DB tournament_picks has 4,419 rows / 50 models / 1,726 resolved. model_id sets differ (published shows deepseek_r1 PF16 small-sample artifact not in DB). Live /audit/ai-tournament.html + /audit/ai_leaderboard.html show a smaller/s…

Resolver workflows return success in <50s but resolve ~nothing. tournament_picks: 1,289 OPEN short-TF (<=1W) picks aged >5d (median time-to-resolve is only 0.99d, so overdue). Throughput collapsed: 1,725 resolved 2026-05-24, then 1, then 0. The backlog alone would push deepseek_v4(43)/cursor_agent(5…

Peer review (VERDICT=REQUEST-CHANGES) of Zoo's uncommitted FOREX consolidation work (M files: alpha_engine/config.py + alpha_engine/non_crypto_policy.py). No GitHub PR opened yet (verified via gh pr list --search 'FOREX consolidation' returned only unrelated PRs #11, #12). THREE BLOCKING FINDINGS: (…

ci-tests.yml fails on main: test_m096_ctf_concentration_cap (4), test_m098_etf_vix_gate (4), test_quality_gates (4), test_pr10_ab_gate (2), test_outcome_resolver_noncrypto (2), test_confluence (2) — subset already fixed by #115.

Stark divergence: ai_tournament_leaderboard.json top model (deepseek_r1) shows PF=16.36, WR=83.3% on n=30 T1. money_ready_verdict.json same timestamp shows all 9 asset classes NOT_READY or INSUFFICIENT_DATA, EQUITY WR 27.5% PF 0.14, CRYPTO WR 37.4% PF 0.89. The tournament picks are not flowing throu…

audit_dashboard/data/pick_summary_stats_48h.json generated_at=2026-05-29T06:38:50Z while sibling dashboard_data.json + money_ready_verdict.json are fresh (2026-05-31T00:02Z). 48h recency panel not refreshing — blocks PR #119 (resolver DESC fix) verification + all 48h Goal #1 freshness claims.

STATE_FILE gitignored (.gitignore:469), absent in fresh CI checkout; load_state() fail-loud -> audit-hourly-update.yml + deploy-competition crash each run. Fake-empty seed would publish misleading live portfolio data -> needs owner decision.

branch-large-file-dup-guard.yml prepends a timestamped alert + commits updates/job-health.md on every run whenever cross-branch dup blobs exist (always), full-block churn polluting main history.

GitHub forces Node 24 default 2026-06-16 (Node 20 removed 2026-09-16). ~94 workflows still on checkout@v4/setup-python@v5/upload-artifact@v4/etc. Zero SHA-pins; tag rewrite safe.

actions_failure_guardian.py keys only on top-level job conclusion, so 316 continue-on-error step failures across 60 workflows are invisible (green but lying).

5-engine NVIDIA NIM panel (Kimi K2.6 + GPT-OSS-120B + GLM-5.1 + Nemotron Super 49B + Mistral Nemotron) unanimously declared COMMODITY the system's #1 alpha, recommending 20-30% allocation. The 3-engine codex/grok/gemini panel (shown the same numbers PLUS leakage signals) classified the same cell DAT…

All asset classes report the same last-pick timestamp to the second. Statistically implausible — strongly suggests this file is auto-generated/simulated rather than computed from real picks. 3/3 swarm REAL.

db_health.json reports 10,501 / 26,945 sampled rows have a >1% pnl discrepancy between stored pnl_pct and recomputed (entry/exit/direction). 12,735 have >0.01% mismatch. Tier: RED. All cohort WR/PF stats built on top of trading_picks.pnl_pct are suspect at this drift level.

data/swarm_picks.json has 38 picks; newest is dated 2026-05-12 (13 days old). Workflow swarm-pick-review.yml runs daily but no longer adds picks — only resolves the existing 38.

tools/top_n_rank_backtest.py read DB_STOCKS_PASSWORD but this host sets DB_PASS_STOCKS. Fell back to default password 'stocks' -> MySQL 1045 Access denied. Also queried asset_class/score columns that don't exist on live trading_picks (uses category/elite_score).

smart_picks_feed pick objects lack the signal_time field. Template logic falls back to age_hours which is computed at dashboard JSON build time. So all rows display the same '1.4h ago' regardless of when the pick actually fired.

fx_algo_performance last 2026-02-09 (all 8 rows 0 picks); at_strategy_symbol_performance (410, only crypto true-PF) snapshot 2026-03-06; algorithm_rolling_perf max calc_date 2026-04-27; algorithm_performance Feb-Mar all-negative. Per-strategy perf not refreshing -> stale per-class edge reads.

Branch docs/eagle4-eagle5-gates-2026-06-02 (carrying PR #447) was fast-forwarded past my 2 EAGLE-4/5 commits (c39d266bfa, 47e3ec4f7c) by concurrent agent activity (PR #455 walkforward-lazy-imports merge). Branch tip moved from c39d266bfa to bda830d7b9 with no EAGLE-4/5 work in the new tip. Rebase te…

Single source system dominates the pick book at 43.6% concentration (HHI=0.232). Should be < 40%.

tools/check_resolver_health.py main() queried forward_test_only counts after conn.close(), surfacing false YELLOW. Fix on branch fix/resolver-health-forward-test-2026-06-02 (PR #457). forward_test columns on at_pick_outcomes were ALTER-applied 2026-06-02 (forward_test_only, forward_validated, _gated…

anti_overfit_audit.json generated_at=2026-05-11 (~20d stale); dashboard_data.json::walkforward.by_class empty, sub-block generated_at=2026-04-15 (6wk). Both run hourly behind '|| echo non-fatal' so they fail silently -> per-strategy DSR table + OOS-divergence detector run on stale data.

refresh_strategy_stats_mysql() landed in PR #91 (audit_trail/mysql_client.py) but is never called from any cron pipeline or workflow. The table remains empty. No feedback loop exists for strategy performance tracking.

During this session agents observed tournament data regressing from 40 models/3,873 picks back to 18 models/478 picks multiple times. Root cause: a peer merge_submissions_to_latest pipeline overwrites the agent-rebuilt JSON with stale data. Auto-restore works but root cause not fixed.

PR #113 was created to rename to .disabled but merge was blocked by branch conflict. The workflow still references STOCKS/competition/forward_test.py which does not exist on origin/main. Every run fails at python invocation step.

After trust_score backfill (32,683 rows + 15,166 floored at trust=2), 15 closed trading_picks rows still have NULL trust_score for strategies that do not appear in any at_signal_outcomes source system. These are orphaned strategies with zero outcome history.

db-backup-email.yml fails: SMTP rejected mail relay (550). Backup itself succeeds but notification fails.

alpha-engine-bond.yml fails: pandas_datareader/fredapi missing + FRED_API_KEY not set. All FRED backends timed out at 30s.

As of 2026-05-31 02:30 UTC, dashboard_data.json generated_at=2026-05-28T21:29Z (age 52.5h). The money-maker-readyv2 freshness gate fails fast >2h. money_ready_verdict.json and pf_registry.json ARE fresh (2026-05-30T23:05Z) so the main verdict pipeline runs, but the dashboard_data.json regenerator (w…

audit_dashboard/data/ma_strategy_leaderboard.json generated_at=2026-05-29T05:27:24Z (~44h stale) with leaderboard=[]. This is the MA Strategy Forward-Tracker v2 output (PR session 2026-05-29). Either the periodic cron is not wired, or tools/ma_strategy_forward_tracker.py is crashing silently.

/audit lists IPO as one of the tracked asset classes but the codebase has zero IPO-specific strategy or pick writer. 1/3 swarm REAL + 2/3 NOISE — added at P2 because the UI advertises the class and the user-facing claim is misleading even if low impact.

Category is NULL/UNKNOWN for 951 active picks (~10% of active set) and 54 closed (35.2% WR). UI can't apply per-class gates to UNKNOWN rows. Cross-class stats undercount these.

ghost_rows audit: 12 cohorts with thousands of identical (asset_class, strategy, symbol, direction, pnl_pct) rows. Top: CRYPTO/quan_engine/MATICUSDT/LONG/pnl=-15.0 with n=20,474 from 1 distinct entry. MEMECOIN/meta_strategy variants make up the next 10. This single cohort alone is dragging quan_engi…

trading_picks.trust_score is NULL on 38,884 of 38,889 closed picks. HC overlay requires trust_score>=4 (CRYPTO) / >=5 (EQUITY). Cited CRYPTO 60.3% N=562 and EQUITY 68.1% N=72 stats unreproducible — only 5 closed picks have a non-NULL trust_score.

Peer claim alleged ~304,000 duplicate rows in ejaguiar1_stocks.at_local_picks. Re-ran tools/at_local_picks_dedup.py (dedup key: symbol+direction+source_system+signal_timestamp) on 2026-05-31 02:49 UTC: table holds 37,763 rows total with 0 duplicate groups. Claim does not reproduce — likely confused …

24,363 SL_HIT rows in at_signal_outcomes: 18,511 have pnl<0 (correct) but 5,852 (24%) have pnl>0 — a stop-loss hit that made money is contradictory. Likely entry_price/exit_price stored with different sign conventions across sources.

FIRING13 report: 198/218 'EQUITY' symbols are mistagged crypto -> real EQUITY n~=20 not 39. Every EQUITY number (n=39 verdict, db_raw 90d PF 5.555 DISPUTED, 48h panel) is unreliable until tagging fixed. Same EXPIRED->WON leakage family as the disputed CRYPTO 78.9%.

at_signal_outcomes EQUITY surged 169→1,424 after PR #121 + live DB backfill. But 0 EQUITY picks meet HC overlay gate (trust>=5). WR=42.1%, PF=0.39 — class is still deeply losing. Goal #1 EQUITY acceptance (>=1000 outcomes) MET but no statistically-valid edge exists yet.

The only WF-VERIFIED equity strategy (62.2% OOS WR on 2-day window) is the new pead_equity, but it never made it past shadow. Meanwhile the broken earnings_drift (0% WR on 92 picks) was active in prod.

Research and backtest evidence is concentrated in cleaner large-cap equity universes, but live EQUITY still carries penny/meme contamination. This distorts both edge claims and gate calibration for the parent class.

The /audit/ueps tab is rendered (n=0/100 disclaimer shown) but no picks have ever been emitted. Composite (Magic Formula x Piotroski x Acquirer's Multiple x SafetyGate) is documented but has no live writer.

code grep found no _run_equity_scanner or similar routing function in production_scanner.py main loop. Strategies (connors_rsi2, quality_compounders, equity_momentum_regime, pead_equity) exist in code but may never be called.

Swarm: deepseek_v4/gpt4o tournament SUPPORT for paper; BAC/JPM/MSFT/NVDA equity picks WEAK. EQUITY class FAIL+INSUFF-N in money_ready (PF ~0.90, WR ~33%, n=33). Production scanner routing unproven.

money_ready ETF INSUFFICIENT_DATA (n=3). Tournament symbols EEM/IWM/GLD show paper WR but cherry-pick ≠ sleeve. Lab etf_verified_dual_momentum is FORWARD_PILOT_ONLY (WF PASS).

etf_dual_momentum, etf_sector_momentum, etf_risk_parity_rotation, etf_faber_tactical, etf_trend_following all allow_without_forward=True. No track record.

CRYPTO directional bug: production emits LONG (33% WR) when SHORT edge is 67% WR

PR #481/#482 bootstrap CI: IS_PF=310.77 on n=67 but pf_lo_95=1.31. One row (id=214622) JUPUSDT pnl_pct=76573 drives gross-profit sum; without it PF collapses. Do not promote to live or forward-pilot until sustained_pf / pf_lo_95>=1.5 on clean sample.

Pick pulse / resolver health flagged top_source_share=0.55 on CRYPTO (live policy-clean cohort sub-T2: PF ~1.14, WR ~43%, n=728 per money_ready_verdict). Concentration gate must hold before any size-up.

Swarm + EAGLE3: tournament SHORT cohort ~67% WR vs LONG ~33% on n≈216. Production scanner historically LONG-heavy. EAGLE-4 flip wired in production_scanner.py but shadow only — policy-clean CRYPTO still NOT_READY PF 0.92.

money_ready_verdict CRYPTO mdd=1.0, cvar_95=-85.1%: fingerprint of mislabeled/never-closed tails inflating loss tail.

policy-clean-net: top_source 55% of CRYPTO book. The only 2 sleeves with PF>1 n>=20 (crypto_liquidity_wick_reversal PF1.55, atr_percentile_gate PF1.10) are both 100% single-source (is_single_source_artifact).

PR #119 changed ORDER BY signal_timestamp ASC to DESC at alpha_engine/active_picks_sync.py:125. Fix landed 2026-05-31T01:20Z. First audit-dashboard.yml hourly run post-fix at 02:10Z will test whether CRYPTO 48h closures emerge from the backlog of 39,706 OPEN CRYPTO picks.

ml_enhanced_INJUSDT_1d_B_lightgbm (n=25 WR 100%), ml_enhanced_DYDXUSDT_15m_D (n=31 WR 96.8%), ml_enhanced_FETUSDT_1d_B (n=25 WR 100%), ml_enhanced_RENDERUSDT_1h_D (n=34 WR 85.3%) on /audit/anti_overfit.html show DSR>=0.9995 as 'publishable confidence' but n is too small for the DSR formula to be rel…

Opencode flagged 1.6M template rows from meta_strategy across MEMECOIN/CRYPTO symbol+direction pairs in backtest_trades. Same root cause as the ghost-rows finding from Qwen's db_health (top 11 ghost cohorts are meta_strategy MEMECOIN). Defer blanket-block until db_health ghost_rows.top_cohorts repop…

Pick-funnel top_edges_per_class found cells like 'copy_trader_intel & LONG' (n=21, PF 1094) and 'conf=0.80-0.85 & ml' (n=42, PF 674) — values that high on tiny samples almost always indicate look-ahead bias in the feature pipeline, not real edge.

edge_decay_heatmap shows quan_engine_scalp at n=4236, WR 37.4%, PF 0.42 — verdict 'dead'. Yet it remains a substantial share of open CRYPTO volume per CLAUDE.md ('18% volume @ PF 0.70 drag elite strategies down').

CR Mean Reversion row shows avg_return_pct=95692.81 on 288 trades. Clear units/aggregation bug; all 8 rows untrustworthy. Do not cite cr_algo_performance.

Investigated 3 suspicious bootstrap CI PASSes (IS_PF>30) flagged in PR #481. Found: - claude_ml_moderate_mut (n=67, IS_PF=310): single row id=214622 JUPUSDT LONG pnl_pct=76573 (entry=0.0002, exit=0.1892, real 945x return). One meme-coin moon dominates the sum. - B_flip_PriceRocMeanReversion (n=1…

DISPUTED banner (Last verified 2026-05-25) cites raw-DB CRYPTO 90d WR 39.4% (n=2001). Live at_raw_picks query 2026-05-31 shows CRYPTO 90d=41.9% (n=7198), 30d=42.6%, 14d=50.9% (recovering), 2d=41.0%. Dispute direction still valid (raw ~42% vs Smart-Picks-claimed 78.9%) but the specific banner figures…

kimi_signal_tracking sign-flip: 142 of 367 total sign-flipped rows = 38.7% (worst single source)

Per benchmark report: forex_carry_momentum, forex_rsi2_mean_reversion, myfxbook_retail_contrarian all losing. Only cta_cross_asset_tsmom SHORT has WR 57.6% but is 93% concentrated in USDJPY — not a diversified edge, just one carry trade.

_FOREX_ALLOWED={cta_cross_asset_tsmom(SHORT),forex_carry}. cta_cross_asset_tsmom has ZERO policy-clean FOREX rows AND anti_overfit rates it n=288 PF 0.254 DSR 0.0 OVERFIT_LIKELY — a proven loser, not 'the sole proven FOREX sleeve' as the code comment claims. The only wired FOREX strategy with a real…

Code review of Zoo's PR#6: the FOREX consolidation gate lives in evaluate_non_crypto_candidate() which has ZERO production callers (live path uses passes_non_crypto_policy() which does not delegate to it) -> the block-all-except-cta_cross_asset_tsmom-SHORT logic never executes. USDJPY concentration …

EAGLE review found the class story is dominated by a few stronger sleeves while the aggregate is dragged down by broad losers. The dashboard does not expose that isolate-the-winner vs kill-the-drag distinction cleanly enough.

alpha_engine/new_strategies/forex_carry.py implements G10 interest-rate differential carry with claimed 55-60% WR / PF 1.2-1.5 but is not registered in non_crypto_policy.NON_CRYPTO_STRATEGY_POLICY so it never emits picks.

Causes 44% SL hit rate vs 12% TP hit (3.7x more stops than targets). After April 2026 widening (TP 0.75%->1.5%, SL 0.5%->0.8%) the situation improved but still asymmetric.

COT publication lag look-ahead bias: cot_positioning PF dropped 4.6 -> 0.51 post-dedup

EAGLE review converged with the existing COT forensic concern: the class story remains unsafe while pre-clean or over-emitted COT history can still dominate class-level PF/WR claims. The page should not treat COMMODITY as trustable until independent-cycle-only stats are canonical.

Benchmark says CRITICAL — cot_positioning at the STRATEGY level is strong (DSR=1.0 per Ring) but at the CLASS level (n=140 closed) numbers are catastrophic because cot_positioning is now BLOCKED per audit benchmark, and remaining cta_cross_asset_tsmom + cta_commodity_momentum_term are losers.

Ring's 2026-05-25 audit says cot_positioning is the SUPREME EDGE (DSR=1.0, WR=86.5%, n=104). audit_benchmark_analysis_2026-05-24 says cot_positioning is BLOCKED and the COT-dedup audit downgraded WR to 5% / PF 0.12 on n=20 post-dedup. Both can't be true.

COMMODITY policy-clean n=9 PF 1.812 collapses to PF 1.15 (drop 1 winner), 0.47 (drop 2 winners). Biggest winner vwap_rsi_confluence +7.3% is symbol SHIBUSDT (a CRYPTO) mislabeled COMMODITY. True clean edge ~break-even, 78% GC=F(gold) concentration. Best wired backtest gold_safe_haven (PF 1.98/n=61/S…

cot_positioning headline (n=104, DSR 1.0, TIER_1) is over-emission: 44 raw emissions of only 6 unique CFTC releases (7.33x). Deduped one-trade-per-release: n=6, WR 33%, cum PnL -,547.57 (NEGATIVE). friction-adjusted MC (cot_step7_friction_adjusted_mc.json n=126) gives DSR 0.0 FAIL / NOT LIVE_ELIGIBL…

Strategy is in code but blocked from production. Either rehab via mutation protocol or formally retire.

audit_benchmark_analysis_2026-05-24.md: BOND class is 0% WR / PF 0.00 / Sharpe -2.465. Only strategy is antigravity_bond with 1 historical pick. Already flagged P0 in Freebuff 2026-05-17.

Bond Emitter workflow has failed daily 2026-05-26..05-30 (~17min each). 5 open bond picks exist in active_picks_bond.json but n_resolved=0 because the emitter/resolution leg is red. bond_connors_rsi2 (PF 1.34/WR 50/Sharpe 2.22/n=201 over 21y) is wired and ready; the blocker is the failing workflow, …

Claims 73% WR but is brand new — needs forward-test data before promotion.

The standalone FUTURES class has near-zero useful activity while real futures exposure is represented under COMMODITY. This makes the /audit taxonomy misleading and blocks honest per-class review.

Both strategies sit in code with BANNED status. Remove from registry to declutter.

alpha_engine/skyrocket_detector.py has the SIDU pattern framework ($0.63->$3.79 example) but is not called from production_scanner.py.

Strategy emits but every pick is rejected at Gate 0 (initial filter). Either fix Gate 0 to allow penny-class scores or move to a class-specific scoring path.

PR landed in audit_trail/universal_pick_resolver.py (_compute_pnl helper @ line 260, wired into 10 call sites). Live DB still shows 1034 WON/LOST rows with pnl_pct=0 from pre-fix round(...,2) collapse. Need targeted backfill: re-run resolution on the 1034 affected pick_ids using new helper. Verified…

Full user-facing guide to the EAGLE-4 admissibility gate (kills 4 noise personas, kills 8 negative-edge class×direction combos, flips CRYPTO LONG→SHORT) and the EAGLE-5 promotion gate (+20% boost for 33 tournament-validated symbols, +15% for 16 promoted personas). 7-section HTML report with ELI5 per…

macd_rsi_m048 passed purged walk-forward validation: 65 trades, 75.4% WR, test PF 3.33, 5/5 folds profitable, negative decay (improves OOS). Best validated strategy in the system.

Jegadeesh-Titman momentum applied to single equities. B-Tier on SPY (Sharpe 1.62, +104%), QQQ (Sharpe 1.63, +165%), GLD (Sharpe 1.18, +128%). Walk-forward validated.

Rotates among SPDR sector ETFs based on 3-month momentum. B-Tier on XLK (Sharpe 1.73, +217%), SPY (Sharpe 1.86, +122%).

Monitors concentration HHI, resolver disputes, per-class health, strategy culling status. Run via python3 verified_strategies/quant_monitor.py.

Turtle Trading-inspired breakout for crypto. B-Tier on BTC (Sharpe 1.91, +371%), ETH (Sharpe 1.69, +574%), BNB (Sharpe 1.57, +358%).

Combines 20-day and 50-day momentum with 200 SMA filter. B-Tier on ETH (Sharpe 1.60, +272%), XRP (Sharpe 1.72, +1169%).

Buy-and-hold mega-cap compounders with SMA trend filter. NVDA +2285% backtest, Sharpe 2.08. GOOGL +733%, AMZN +441%, JPM +336%.

Consolidated findings from Claude Opus 4.7, minimax-m3-free, Mimo v2.5 Pro, Grok. Identified AI tournament edge matrix, directional biases, best symbols, best personas.

Replace mutation_framework.compute_pf with proper PF formula + reject 999 fallback inflation

Tournament data: SHORT 67% WR / +3.74% avg PnL vs LONG 33% WR / -0.49% avg. Flip production scanner to SHORT-only for CRYPTO.

Fix MC null hypothesis: bootstrap-with-replacement -> block bootstrap

DSR n_trials: read real cumulative hypothesis count from reports/hypothesis_registry.json

Land block-bootstrap helper in verified_strategies/strategy_verification_engine.py (file now on main via PR #455)

Ship check_resolver_health.py fix (forward_test before close), tools/ensure_forward_test_outcome_columns.py idempotent ALTER, and stale-open batch-resolve documentation.

is_single_source_artifact currently advisory. Make it a hard reject (or require >=3 sources) before any sizing. Kills the 2 false-positive crypto sleeves until de-concentrated.

Morning checklist: run_eagle_suite.py (--skip-swarm on laptop), verify money_ready_verdict, FTP strategy_admissibility.json + pilot_forward_dashboard.json, zero sizing on NOT_READY.

All green pick_funnel cells carry discovery disclaimer; cross-link strategy_admissibility.json and money_ready_verdict. Prevents tournament/lab numbers sizing production.

From reports/2026-05-26_forward_validator_freeze_diagnosis.md + money maker v2 unblock plan. Current incidents.html / DB entry misattributes the 270h freeze. Accurate: outcome-resolver.yml failing on missing closed_picks.json in git add since 2026-05-25 19:54. This is the real blocker for signal_out…

From reports/2026-05-26_forward_validator_freeze_diagnosis.md (cross-ref money maker v2 unblock plan). The 'forward validator frozen 270h' claim is misdirected. Real freeze is in outcome-resolver workflow: fatal pathspec 'alpha_engine/data/closed_picks.json'. Last success 2026-05-25 18:51, failures …

From daily_ideas_edge_sweep_2026_05_17.md + synthesis + this Daily Ideas subagent. Nightly snapshot of information_schema to JSON + automated diff against version-controlled baseline in schemas/. Currently no such workflow despite high consensus. Integrates with existing db-freshness guardian and cr…

From daily_ideas_edge_sweep_2026_05_17.md + synthesis + this Daily Ideas subagent. Remove '|| echo "non-fatal"' in mysql-trading-sync.yml line 43 (and similar) that swallows DB failures. 4+ agent consensus; 5-min fix. Aligns with existing db-freshness guardian.

From daily_ideas_edge_sweep_2026_05_17.md + synthesis + this Daily Ideas subagent. Nightly snapshot of information_schema to JSON + automated diff against version-controlled baseline. Currently no such workflow despite high consensus. Integrates with existing db-freshness and cross-DB audit.

From AUDIT_DASHBOARD_BURIED_WINNERS_2026-05-29.md + this subagent. Every leaderboard entry appears exactly twice (source="" + real); inflates counts. alpha_engine aggregate losing (−5.56, 37.1% WR) while sub-strats (mega_mutation etc.) win big. Dedup by strategy+source + surface buried + per-strateg…

From 2026-05-29-COMPREHENSIVE-QUANT-ANALYSIS.md + this subagent (Tier 1 recs #3/#4). incidents.html timestamps broken; no audit trail for enhancement plans/dates. Add schema columns + population + generator fixes for credible /audit/incidents.html as Goal #1 task board.

Add post-render job + TRIAGE DASHBOARD section on incidents.html showing P0 counts, age>7d escalation, and OPEN->RESOLVED->OPEN regressions. From 2026-05-25_incidents_triage_consult_codex + grok. Makes nightly incidents page credible for Goal #1 work.

Ungrounded 5-engine panel called COMMODITY #1 alpha (PF 4.64 from 85% CT=F leakage); grounded panel correctly flagged DATA_QUALITY_LEAKAGE. From 2026-05-25_multi_ai_panel_meta_review. Mandate leakage signals + hypothesis_registry rejected entries in every consult prompt.

Nightly seed_incidents_enhancements.py re-inserts destroy triage state and create false 'new' P0 churn. From 2026-05-25_incidents_triage_consult_codex.md + gemini. Add deterministic finding_key (asset_class + issue_type + normalized_title) + INSERT ... ON DUPLICATE KEY that only touches evidence/upd…

Roo's NIM panel session (2026-05-25) reports moderate_confidence (n=455) and low_confidence (n=193) buckets went 0-for-648 over the 6-day window 2026-05-16..21, destroying -825% PnL, while 300 gated picks generated +994%. If real this is the single highest-leverage filter in the system. 0-for-455 is…

Qwen step-2: assert no test-period features were populated during training. Check columns NaN pre-cutoff AND populated post-cutoff as red-flag for forward-fill leakage. Raise DataLeakageError on detection inside every fold iteration.

Opencode P0 DRAFT. UPDATE pass that re-computes status from pnl_pct for any (status='WON', pnl_pct<0) or (status='LOST', pnl_pct>0) row. Sign-coherence guard already stops NEW contradictions; this backfills the historical 2,531+ WON rows with negative PnL flagged in db_health.json::won_pnl_contradic…

Cheapest single fix per persona-survey — addresses 7 personas / ~470 picks. ~30% of picks fire in the wrong regime today.

ai_tournament_leaderboard.json tags 67-91% WR models as tier=T1/T2 with NO per-row disputed/artifact/caveat field. Per memory ai-tournament-wr-artifact-2026-06-03, these WRs are suspected single-snapshot resolver artifacts (fix = intrabar OHLC replay). The DISPUTED banner is HTML-only — non-browser …

at_strategy_stats covers only CRYPTO/MEMECOIN (+thin EQUITY). FUTURES (4094 raw picks, 246/14d), ETF (421), BOND (90), COMMODITY (5723 outcomes), PENNY (494) have ZERO strategy-level perf rows — only raw/outcome ledgers. Can't evaluate per-class edge without tracking. Build per-class strategy stats …

SL/TP audit using intrabar OHLC replay (EAGLE_JUNE2 §7.1 #3)

Pre-register, data validation, purged walk-forward, cost model, DSR/PBO, block bootstrap, regime robustness, concentration check, forward validation, promotion decision.

Daily verified pilot loop (PR #452 MERGED): ETF dual_momentum + crypto WF forward stats + faber forward stats

Replace bootstrap-with-replacement with moving-block bootstrap (Kunsch 1989) for Monte Carlo null hypothesis testing. Preserves serial correlation, trend/momentum structure, and volatility clustering.

Implement Probability of Backtest Overfitting per Bailey & Lopez de Prado (2015) using parameter grid permutations. 5 perturbation types: window shift, threshold flip, holding period, trailing stop, volatility scaling.

ETF Dual Momentum (Antonacci 2013) is only lab strategy passing Tier-2 gates: PF 1.60, n=104, Sharpe 1.91. First candidate for paper trading pilot.

Second in-pipeline gate (6f2.6 in production_scanner.py) that BOOSTS confidence for 33 tournament-validated symbols and 16 personas that have demonstrated >=60% win rates. Multiplicative boost: +20% on whitelisted symbol (_EAGLE5_SYMBOL_BOOST=1.20), +15% on whitelisted persona (_EAGLE5_PERSONA_BOOST…

First in-pipeline gate (6f2.5 in production_scanner.py) that operates on the 4 noise personas + 8 negative-edge class x direction combos identified in the AI tournament analysis (5,492 picks, 3,692 resolved). Three actions: (1) KILL 4 NOISE personas — momentum_scalp (28% WR), breakout_scanner (28% W…

Wire promotion_gate.is_admissible_for_production() into production_scanner emission path

Add current_price + unrealized_pnl_pct columns to PF_POSITION + daily mark step

Real k-fold PBO using 8-fold purged WF (already coded at rigorous_backtest_harness.py:249-320 but never called)

Symbol whitelist: restrict production emission to AI-tournament-proven symbols (BAC/JPM/MSFT 64-100% WR; KULR/RGTI/ASTS/RKLB 83-100% WR for penny)

Cron/local: run_eagle_suite.py then run_verified_pilots_daily.py (ETF dual momentum pilot, crypto_wf/faber/etf forward stats, pilot_forward_dashboard.json). GitHub workflow verified-pilot-daily.yml 06:15 UTC.

After first 5×500 batch (~100k→~3.8k OPEN), schedule additional resolve_stale_open_picks.py runs until OPEN count matches hold-window expectations per asset class.

Add cross-section multiple-testing correction (Bonferroni or BH-FDR q=0.10) as a hard pre-gate before DSR/SPA in money_ready_verdict. Closes the false-Tier1-PASS hole.

Final production pipeline gate (6f2.7 in production_scanner.py) sitting on top of EAGLE-4 (6f2.5) and EAGLE-5 (6f2.6). Kills picks whose strategies fail any of: (a) DSR noise filter — strategy NOT in the 27 dead strategies from tools/deflated_sharpe_results.json (sr_haircut=2.5034 for 164 strategies…

PR #458 and #457 merged on main 2026-06-02. walkforward_suite pilot PASS etf_dual_momentum OOS PF 2.75. #447/#444/#461 still open.

Weekly SQL from EAGLE doc Part E: CRYPTO/FOREX EXPIRED wins, duplicate signal_ts groups, depromote battleground/regime_terminal concentration emitters.

Re-run walkforward_suite on 50webs when source available; strategy_admit per sleeve; Bonferroni α_adj from hypothesis_registry.json (M-107).

Fleet investigation: the bottleneck is NOT strategy discovery or Monte-Carlo (validation is over-served, runs hourly) — it is live signal VOLUME. Multiple PROVEN backtest edges exist but emit ~0 live picks: ETF etf_cross_sectional_momentum (PF 2.05/WR 70/Sharpe 0.97, wired+ENABLED, just young); EQUI…

Tiered continue-on-error policy: manifest grandfathers 38 existing silent maskers; lint_workflow_masking.py --fail-on-new blocks only NEW silent maskers at PR time (zero new red X). PR-only, never on push-to-main.

Extend failure-guardian to surface green-job/failed-step masking (the 316-coe blind spot) via the /runs/{id}/jobs API, bounded for quota, report + Discord.

From reports/CYCLE_12_STRATEGY_HUNT_2026-05-29.md: 'Key Discovery: **ADX_Trend is a universal winner across ALL asset classes** — the first strategy to show PF>1.5 on EQUITY, ETF, COMMODITY, FOREX, and CRYPTO simultaneously.' Table with specific Tier 1: 'ADX_Trend | NVDA | EQUITY | 5.17 | 80.0% | 10…

From reports/STRATEGY_ROADMAP_COMPREHENSIVE_2026-05-29.md: Strategies with DSR/PBO computed | 2. Average PBO (all sized) | 0.613. Root Cause: High PBO (0.3-0.7) indicates most strategies were data-mined. The solution: fewer parameters + economic rationale + purged walk-forward. The 7 New Strategy De…

From daily_ideas_synthesis_2026-05-16.md + edge_sweep + this Daily Ideas subagent. 22/38 swarm picks are tier=single (1/1 vote) with no backing. Operational risk; backfill + enforce tier-gate so only verifiable high-quality singles get promoted to TV-eligible/Smart Picks.

From daily_ideas_synthesis_2026-05-16.md + edge_sweep + KimiCode + this Daily Ideas subagent (HIGH, fixes anti-predictive confidence inversion blocking all WR/PF claims). Create at_confidence_calibration (bucket-level drift tracking + auto-quarantine) + tools/mysql_prediction_anomaly_scanner.py (inv…

From daily_ideas_synthesis_2026-05-16.md + edge_sweep + this Daily Ideas subagent review (multi-agent consensus). 5-gate exec-time reject layer (portfolio concentration + correlation + regime + thin-sample + other) proposed for shadow to block live sizing on correlated/thin/quarantined signals. No p…

From AUDIT_DASHBOARD_BURIED_WINNERS_2026-05-29.md + this subagent review. Smart Picks only has 4 picks despite 16+ strategies with forward PF>2.0. mega_mutation PF4.39 n=110; kimi_riseoftheclaw scouts PF>2.5 with 0 active. Add per-strategy filtering, promote buried high-PF to active sizing, add heal…

From 2026-05-27_pr_set_plan_and_blockers.md + AUTONOMOUS_STRATEGY_HUNT_CYCLES + this subagent. No current gate; small-sample luck survives. Add tools/permutation_test.py + wire to quality_gates/promotion: require real PF in 99th percentile of noise. Proven in hunt (91.5% profitable, ensemble 92% sig…

From daily_ideas_synthesis_2026-05-16.md (KimiCode + synthesis). Inverted confidence, direction conflicts, silent-dead strategies, drift tracking. Directly improves Goal #1 data integrity and incidents visibility.

Create a non-admission-changing observability lane that records picks rejected by gates or quarantine rules but later resolved positively. This turns hidden false negatives into measurable backlog without weakening live safety gates on day one.

Current streak logic influences scoring but does not create a controlled exemption path. Add a time-boxed, per-sleeve exemption contract so repeated clean winners can earn temporary gate relief without silently changing the system.

10-model panel (NVIDIA DeepSeek-R1-32B + kimi-k2.6 + CF qwen3-30b/qwq-32b/llama-3.3-70b/llama-4-scout/nemotron-3-120b/gpt-oss-120b/glm-4.7-flash/mistral-small-24b) converged on 8 core scoring dimensions: Sharpe(20%)+Sortino(20%)+MaxDD(20%)+n(10%)+Bonf-sig(10%)+OOS-ratio(10%)+cost-burden(5%)+bias-aud…

Qwen gate-10: replay 2020-03-09 through 2020-03-23 (flash crash) on candidate strategy. Fail deploy if drawdown >30%. Catches strategies that work in normal regime but blow up in correlation-to-1 events. Add stress_test_2020_march() helper, run once per candidate in CI.

Qwen pitfall #4: 4 'different' US-tech holdings = 1 sector. Use scipy.cluster.hierarchy.linkage(method=ward) on correlation distance matrix, cap risk per cluster <=20%. Replaces naive sector-string caps in portfolio_construction.

Qwen gate-9: shift each key parameter +-20% and re-run OOS. Fail deploy if any single shift causes >15% PnL degradation. Detects fragile parameter overfits. Add _run_parameter_sensitivity_analysis() to StabilityReport.

Agent B (c9ba1589) added Bonferroni to top_edges.py. KEY FINDING: CRYPTO's 5 apparent edges DO NOT survive correction across 673-cell space; only COMMODITY produces robust edges. Currently live score_pick / passes_smart_gate uses raw thresholds. Wire wr_z + bonferroni_pass fields from top_edges outp…

Qwen step-4: shuffle outcome labels N=100, re-run walk-forward, derive empirical permutation p-value on mean OOS Sharpe. Required gate: p<0.10 before shadow->probation promotion. Add _shuffle_and_revalidate() helper, return permutation_p_value field in WalkForwardResult.

Current portfolio constructor likely uses sample covariance which fails on regime breaks (Mar 2020 correlations spike to ~1 and equal-risk allocation blows up). Switch to exponentially-weighted covariance with Ledoit-Wolf shrinkage; recompute on a rolling basis. Required input for the planned CVaR o…

Ring's quant-workflow spec: a systematic walk-forward runner with purged k-fold + embargo to prevent leakage. Currently no such runner exists in repo; DSR sidecar is the only overfit guard. Complements the kilocode/nemotron purged-k-fold proposal but is more specific (30d purge, 5d embargo, >=8 fold…

Per Qwen framework: OOS Sharpe>=0.5, max DD<20%, WR binomial p<0.05, Sharpe CoV<0.5, edge-decay slope>-0.05, permutation p<0.10, net-Sharpe-after-costs>0.30, zero leakage flags, parameter sweep stability <15% PnL loss, stress-test (Mar 2020) DD<=30%. Failures block promotion or force reduced sizing.

Kilocode shipped a reference pseudo-code: 50%/50% blend of (a) rolling-z 20d log-return momentum, (b) rolling-z macro-surprise series (CPI/PMI/yield-curve from FRED). Outputs a rank-able composite score. Closes the long-standing 'FRED data wired but not used in pick gen' gap from HEDGE_FUND_GAP_ANAL…

Opencode found 206 files in alpha_engine/baby_strategies/, ZERO wired to production. Massive untapped pipeline. Surface a batch DSR runner (anti_overfit_audit_sidecar.py over baby_strategies/*) to find DSR-real candidates and promote them per the Wire-Up Rule.

Backfill from strategy registry so HC overlay claims become reproducible. Unblocks all 'closed-book edge' callouts on the page.

After each AI-tournament submission, re-prompt the model: 'Are these picks based on cited live market data or speculation? Mark each.' Store flag in tournament_pick_research.research_basis.

docs/EXEC_PLAN_06 plans Day-75 promotion of vrp_harvest_v1 to PROMOTED_STRATEGIES. Independent 2-source review (Claude + DeepSeek-Chat-Direct via local LiteLLM :4000) found 11 critical bugs in verified_strategies/vol_risk_premium_harvest.py: (1) variance-units-to-returns dimensional bug, (2) ^VIX in…

Per KTD-Fin (arXiv 2605.28359): LLM-agent backtest edge fails via knowledge leakage (recent window in model cutoff = memorization) + return attribution (beta/style not alpha). Our ai-tournament deepseek_v4 PF3.46 is on a recent named-ticker window -> doubly suspect.

Standalone verification engine combining block-bootstrap MC null, PBO, DSR, purged walk-forward, regime analysis, cost model, and tier assignment. Single source of truth.

Backlog for EAGLE-6 v2: add three more statistical gates on top of v1's DSR+n+HHI to harden the final admissibility check. (a) PBO<0.5 — Combinatorially-Purged Cross-Validation Probability of Backtest Overfitting. Blocked: tools/cpcv_pbo_results.json not yet generated. (b) Walk-forward OOS PF >= 0.8…

30-day shadow paper pipeline for ETF Verified Dual Momentum (first promotion candidate)

Tournament (ai-tournament.html, deepseek_v4 PF3.46 paper) and production money_ready use different labels/costs. Unify: one purged-embargoed walk-forward harness (30d purge+10d embargo) as the only promotion path.

Final consolidated 162-line spec from Grok (3-round swarm review + DB investigation + Quant 4-phase vetting). Unified 5-sub-score framework with class-specific factors for EQUITY/ETF/CRYPTO/FOREX/COMMODITY/FUTURES/BOND + strict high-risk bucket for PENNY/Cheap/IPOs. Grounded in real DB findings (top…

Replace current rank-and-fill logic with constrained min-vol optimiser that includes (a) CVaR constraint, (b) sector exposure caps, (c) turnover penalty inside the objective. Operationalises Qwen's pitfall #2 (transaction cost feedback) and pitfall #4 (hidden sector concentration).

Kilocode/Nemotron 2026-05-25 session proposes: expanding-window (T0=250d / 21d test), 5-fold purged TS-CV with 5-day gap, Bayesian-shrunk Sharpe + bootstrap CI, Ljung-Box on residuals, Monte-Carlo return-shuffle (>95th percentile), regime-split via HMM. Currently only DSR sidecar enforces overfit gu…

Build tools/ai_tournament/consensus_tier_algorithm.py: for each (asset_class, feature_concept), take median weight across all models in tournament_rating_algorithms. Features with >=2-model consensus seed alpha_engine/score_v3.py.

Qwen 2026-05-25 framework: (1) PipelineConfig, (2) ingestion, (3) momentum+macro features, (4) regime detection, (5) signal generation, (6) inverse-vol + capped-Kelly sizing, (7) CVaR-aware portfolio constructor + execution queue. Modular swap-in for any stage. Pairs with the 8-pitfall avoidance + 1…

money_ready_verdict.classes.<CLASS>.wr is a decimal (e.g. CRYPTO=0.3613). pf_registry.by_asset_class_policy_clean_net[<CLASS>].win_rate_pct is a percent (e.g. CRYPTO=36.129). Consumers reading the wrong field show off-by-100x WR. Either unify (rename one, add converter, or expose both with explicit …

Category-taxonomy normalization: collapse stock/stocks/penny/pennystock to EQUITY + symbol-pattern auto-tag for NULL/UNKNOWN

Captures USD regime shifts across major pairs. B-Tier on USDJPY (Sharpe 2.02, +130%), C-Tier on EURUSD/GBPUSD.

Agents can now run verify, monitor, mutation, deploy, eagle, litellm, swarm, consult without human babysitting. Auto-injects --yes for tools that support it.

Full HTML report at updates/eagle2-quant-review-2026-06-02.html. Root causes, best picks, walk-forward validated strategies, short/long-term plans.

PR #457 OPEN: resolver health forward_test_tag_awareness fix + idempotent DB ALTER helper

Deflated Sharpe Ratio n_trials must reflect actual strategy variants tested. Load from hypothesis_registry.json with conservative floor of 500.

Compute CAGR + Sortino in compute_metrics (currently null on portfolio drill pages)

Roo's session reports: kimi_signal_tracking 168 picks WR 53.6% +257.34% (best source by total PnL); aggregated_picks 58 picks WR 74.1% +111.02% (underpowered but interesting). Run the same per-source rollup over a longer window (30d/90d) to test if these are persistent edges or 6-day noise. Apply th…

When ex-ante vol estimate is too optimistic (predicted 8% vs realised 14%), position sizes are too large and drawdowns blow through risk budget. Need a daily job that logs predicted vs realised vol per asset class and auto-widens sizing buffers when realised > predicted for N consecutive days.

One-line addition in dashboard_generator.py. Stops the 'all picks show 1.4h ago' misleading display.

Investigate TP_HIT_REPLAY regression in audit_trail/outcome_resolver.py bar_replay()

Any sleeve clearing the harness runs at <=0.5% live for 4-8wk; promote to real capital only after 2 consecutive 4-week windows within +-10% of backtest PF.

Qwen step-4: np.polyfit(range,rolling_sharpe.dropna(),1)[0] slope. If slope<-0.05, flag as 'edge appears to decay, possible regime overfit'. Add to StabilityReport, deploy gate slope > -0.05.

Qwen pitfall #6: shipped as vol-calibration tracker (f54ee2ed). This entry tracks the WIRE-UP into the active deploy gate: surface realized/predicted on /audit, gate strategy promotion on trailing-3mo median<=1.3.

Calendar-rebalancing during flash crashes moves the market against you. Replace 'rebalance every Mon' with 'rebalance when drift exceeds 1.5x expected'. Use VWAP / iceberg slicing on any single fill > 0.5% ADV to limit market impact. Reduces rebalance turnover by ~40% historically.

Currently the AI tournament locks symbols to 2026-05-19 snapshot. Widen to S&P 500 + active /audit picks per class so cross-system comparison is apples-to-apples.

Show why each symbol scanned but not picked was rejected (which gate killed it). Pick-funnel automation already extracts this; needs UI surface beyond /audit/pick_funnel.html.

deepseek_v4 #1 tournament leaderboard (n=208, PF~3.5 paper). Bridge tournament picks through admissibility — separate universe from at_raw_picks until virtual forward book matches production gates.

Grafana-style panel per EAGLE2 §4.6: per-class PF/WR/MDD, portfolio HHI <0.20 target, resolver dispute alerts, admit pipeline latency ≤5 min/sleeve.

From reports/2026-05-27_quick_wins_from_90day_plans.md + equity_vix_regime_breakthrough_20260513.md + consolidated_plan_2026-05-29.md. Backtest lift: VIX<22 PF 4.55 / MDD 16.8% vs baseline 2.82/24% on large-cap universe. Single production caller needed per Wire-Up Rule. Goal #1 critical for EQUITY e…

The main EQUITY sleeve should reflect the clean large-cap / regime-controlled strategy set, while penny/meme names live in a separate research-only bucket. This improves both reporting honesty and future gate calibration.

Only WF-VERIFIED equity strategy (62.2% OOS WR). Currently dormant.

Pure-consensus (no backtest) stability-tilted basket from 6-model panel. Top: MSFT(6/6), BRKB(5/6), SGOV, VOO; AVOID INTC, NVDA divisive. Documented in ejaguiar1_backtests.eagle2_consensus_picks + eagle2_methodology.

48mo clean-bar mega-cap momentum: PF3.53 Sharpe1.65 but #111 attribution vs SPY FAILS (alpha t=1.98<2.0, beta1.13=amplified equity beta) + survivorship bias (hindsight universe). Not promoted. Needs point-in-time universe to be a real test.

etf_cross_sectional_momentum (Jegadeesh-Titman 12-1 top-3, long-only) backtests PF 2.05/WR 70.5%/Sharpe 0.97/MDD 16.1%, slippage-robust (PF holds 1.99-2.05 across 5 friction scenarios), n=122mo 2015-2026. Already registered in alpha_engine/etf_scanner.py and ETF_CS_MOM_ENABLED defaults ON. Only gap …

PR-merged backtest: 48mo walk-forward dual-momentum PF3.57 Sharpe1.62 MDD-12.4%; clears #111 attribution vs SPY (t=2.36 IR0.37 beta0.34=real alpha), bootstrap PF CI[1.64,9.69], cost-robust 20bps. FIRST sleeve to clear the gate-stack.

Lab WF OOS PASS (PF ~1.21); production merge OFF. Daily etf_dual_momentum_pilot.py logs virtual signals to etf_dual_momentum_paper_log.jsonl. First OPEN 2026-06-02 XLK; n_closed=0; promotion_ready=false until n≥100, PF≥1.5, WR≥50% (shadow checkpoint n≥30).

EAGLE review identified ETF sector rotation plus VIX gating as the cleanest underused edge in the current repo. Existing mixed ETF sources dilute that cleaner regime-aware strategy story.

Only multi-class lab Tier-2 pass (PF 1.60, n=104, WF OOS PASS). Admit FORWARD_PILOT_ONLY. Gate: etf_forward_stats.promotion_ready, MDD<15%, live PF ≥0.9× backtest, ≤0.5% shadow capital.

ETF is best lab edge / low concentration but n=3 live. Build absolute+relative dual-momentum (SPY/EFA/AGG/GLD), monthly rebalance, 24mo walk-forward.

Roo's session reports that regime_adaptive x ETF is the only persona-asset pair passing all statistical gates (binomial significance + positive PnL + positive Sharpe), Wilson CI 49.7-91.8%. Cross-reference with the prior 30d ETF PF=3.88 'STRONG RECENT' regime-shift thesis. If confirmed, this is the …

spotgamma/unusualwhales feeds. gamma_raid currently narrates gamma without consuming it; persona is already at WR 67.9% — real data should lift it further.

Promote macd_rsi_m048 to PROMOTED_STRATEGIES after 30d shadow

From reports/asset_class_90day_plan_CRYPTO_2026-05-15.md + 2026-05-27_quick_wins + consolidated_plan_2026-05-29.md + EAGLE sessions. Memory n>1000 shows clear WR edge in those hours. Env-gated (CRYPTO_UTC_HOUR_FILTER). Highest-leverage low-effort CRYPTO item from 90-day plans.

WF: vwap_reversion PASS (PF 1.32, n=516 OOS); bollinger_mr PASS (PF 1.67). Hyro pilot sleeves gated by CRYPTO_VERIFIED_* env. Donchian FAIL — do not enable.

atr_percentile_gate has a winning entry (58.6% WR, n=29 policy-clean) but losers exceed winners (avgW +0.42% vs avgL -0.54%, |W/L|=0.78), capping PF at 1.10. Loss-cap counterfactual: SL @ -0.5% -> PF 1.47; SL @ -0.4% -> PF 1.67 (crosses Tier-2). Winsorization is upper-bound; needs price-path backtes…

Closest CRYPTO money-ready sleeve: n=43, 58.1% WR, PF 1.498 policy-clean. 40% of exits are TIME (timeout) not TP/SL (TP:10 SL:16 TIME:17) -> exit timing mistuned. Loss-cap counterfactual: SL @ -0.5% -> PF 2.47; SL @ -0.4% -> PF 2.96. Needs n>=100 to reach money-ready. [VALIDATED 2026-05-31: price-pa…

From daily_ideas_edge_sweep_2026_05_17.md + synthesis + this Daily Ideas subagent. Wire meta_labeler.py into passes_active_gate; CRITICAL drop of forward_wr (leak) + confidence (inverted) from features. Partial shipped (code exists) but production gate wiring OPEN. +8-15pp WR projected on filtered s…

genome mega_mutation_MACD_RSI: n=210, 68.6% WR, PF 4.276, diversified (JUPUSDT 21.9%), balanced L/S (107/103), clean exits (TP 141 SL 65). EXCLUDED from policy-clean (collapses to n=1) because genome is a genetic-evolution BACKTEST engine -> likely in-sample overfit. If it survives strict OOS forwar…

Top recommendation from persona-improvement survey: addresses CLAUDE.md CRYPTO PF 1.25 -> T2 PF 1.5 gap. Covers 6 crypto personas (~453 picks).

Swarm SUPPORT: deepseek_v4/gpt4o tournament sleeves; CRYPTO SHORT ~67% vs LONG ~33% on reviewed cohort. Capital decision NO-GO on production /audit sizing until money_ready clears.

Both JSON files generated at 2026-06-05T05:33Z but report different FOREX figures: money_ready.classes.FOREX = {n_resolved:23, wr:0.2174, pf:10.80} vs pf_registry.by_asset_class_policy_clean_net[FOREX] = {n:22, win_rate_pct:22.73, pf:11.22}. 1-pick + 0.42 PF gap means the two pipelines have divergen…

From daily_ideas_synthesis_2026-05-16.md + edge_sweep + this Daily Ideas subagent (4+ agent consensus, P0/CRITICAL for FOREX rescue from FAIL PF 0.27). Scaffold tools/research/forex_carry.py (monthly rebalance, AQR 30yr Sharpe 0.7-0.9); FRED_API_KEY present. Hard-disable in place until shipped + n>=…

Implementation already in repo (alpha_engine/new_strategies/forex_carry.py) with G10 rate differential, claimed 55-60% WR. Only missing the allowlist entry.

48mo clean-bar FX trend: attribution vs UUP PASSES (alpha t=2.13, beta -0.61 short-USD), but Sharpe 0.42 + bootstrap PF CI[0.63,3.36] lower<1. Real but small/unstable alpha. NOT promoted. FXE fetch-failed (4-ccy universe).

The class should only advertise edge using independent-cycle-aware COT accounting. This enhancement formalizes the honest source-of-truth rule and blocks stale, over-emitted history from defining the class story.

Top-2 data-feed investment from survey. Addresses sub-floor FOREX/COMMODITY classes by giving cta_trend / supply_demand / inventory_cycle real positioning data instead of inferring from price.

Opencode P2 PASSIVE. Gates the only currently-DSR-verified single-class deviation candidate (cot_positioning + CT=F). Steps 1-5 ~6h active work; Step 6 = 4-week paper pilot (currently SHADOW); Step 7 = Monte Carlo risk-of-ruin sim.

Clean-bar 48mo TSMOM: PF1.69 Sharpe0.67 MDD-33.8%. Fails #111 (alpha t=0.84, beta0.72 vs DBC) + bootstrap PF CI[0.78,3.93]. Mostly commodity beta, no alpha. Gate-stack correctly rejects (vs ETF which passed).

From 90-day BOND plan + Cycle 17 hunt (PF 2.11 Tier 1) + consolidated_plan_2026-05-29.md. BOND class was 0% WR n=9; now has proven edge but still not emitting at scale. FRED key + wiring = high leverage for Goal #1.

alpha_engine/bond_scanner.py exists with yield_momentum, duration_rotation, mean_reversion; not currently wired into production_scanner main loop. Universe of 14 symbols ready at config.py:721. Wiring should lift BOND n from 18 to 50+ within 2 weeks.

From daily_ideas_synthesis_2026-05-16.md + edge_sweep + this Daily Ideas subagent (P1 for BOND/ETF lifts from thin/INSUFF-N). Expand bond_scanner.py beyond TLT/HYG to full 14-symbol; add relative-strength + macro overlay (VIX/YC/DXY) to ETF sector rotation to push PF 1.33→1.5.

From CYCLE_17_FOREX_BOND_BREAKTHROUGH_2026-05-29.md + this subagent. 8 Tier 1/2 BOND candidates discovered; mean-reversion now viable on bond futures (previously unviable). Register + wire with n>=10 / p<0.05 gates.

From daily_ideas_synthesis_2026-05-16.md + daily_ideas_edge_sweep_2026_05_17.md (multiple agents consensus). BOND class thin (n=11). Wire the existing bond_scanner.py fully and expand universe per 90-day + daily ideas recommendations. Goal #1 blocker.

BOND class has zero live samples. Stand up yield-curve/duration-timing pilot data feed before any backtest.

Use new_strategies/tsmom.py framework to trade the 10Y-2Y curve via TLT vs IEF. BOND class currently has only bond_connors_rsi2 (new, no track record).

48mo clean-bar duration-timing: PF1.08 Sharpe0.10, #111 attribution vs AGG NEGATIVE alpha (t=-0.63 beta0.50) + bootstrap[0.53,2.41]. 2022-25 rate regime. Closes BONDS#7 data-gap. Only ETF dual-momentum clears the gate-stack of 5 archetypes.

A unified futures taxonomy would stop the page from showing a nearly empty standalone FUTURES class while real futures exposure is discussed elsewhere. This is primarily a reporting/trust fix before it is a strategy expansion.

Use cta_commodity_momentum_term framework. Captures contango/backwardation premium — proven hedge-fund recipe.

Penny stocks have only one (unwired) strategy. Build float-squeeze + volume-breakout signal using existing SIDU pattern code.

audit_dashboard/ai-tournament.html calcModelStats() L517-521 computes m.score = m.wr * m.pf with m.wr_ci_lo = m.wr and m.wr_ci_hi = 1 — i.e. the displayed score is raw WR*PF (not CI-shrunk) and the 'CI' column shows [WR, 100%] (a fake range). Meanwhile the page header (L190), explainer (L201, L207),…

evidence: Resolved via commit df65641d8 (fix/leaderboard-real-ci → main): added closed-form Wilson WR_lo/hi + bootstrap PF_lo (B=500, seeded LCG) per deepseek swarm consult Option B; score now uses real CI valu

PEER-VERIFIED FINDING (2026-05-31 / claude-opus-4-7-desktop) SYMPTOM ------- Filtering at_raw_picks by `pnl_pct IS NOT NULL` returns one-sided WR for many strategies that pf_registry (policy-clean) reports as balanced. - crypto_liquidity_wick_reversal_v1: at_raw_picks shows 59 WON / 0 LOST, but pf_…

evidence: Query: SELECT strategy, SUM(status='WON'), SUM(status='LOST'), SUM(status='OPEN'), COUNT(*) FROM at_raw_picks GROUP BY strategy HAVING (won>=10 AND lost=0) OR (lost>=10 AND won=0) → 41 rows. NULL-by-s

Three open PRs are related: #75 (per-class FINDING tables) is subsumed by #76 (_UNSET sentinel + CLI), which is subsumed by #78 (CLI + renderer + test_cli_finding.py). #78 is the superset. However #78 includes 2 stale data JSONs (ai_tournament_leaderboard.json + ai_tournament_model_summary.json from…

evidence: Verified via gh pr diff 75/76/78 + git show origin/main:tools/audit_pick_funnel/cli_track.py (does NOT contain _UNSET sentinel). CI run 26672026454 test 3.11 FAILED. All 4 hold-comments from Opus 4.7

Pipeline run 26641993453 logs show repeated auth_error (403) Key limit exceeded across every OpenRouter-routed model: claude_haiku_4_5, command_a, minimax_m2_5, kimi_k2_6, qwen3_5_plus, qwen3_6_max. Operational fix: top up OpenRouter credits OR rotate to fresh OR API key in GH secret OPENROUTER_API_…

evidence: GH Actions run 26641993453 job 78522642787 log; lines like '[API] anthropic/claude-haiku-4.5 auth_error (403): Key limit exceeded (total limit)'

tools/portfolios/meta_effectiveness.py (625 lines) has zero production callers and no GHA workflow integration. 11 tests all pass. Math is solid (Sharpe, Sortino, MaxDD, CAGR, Calmar, attribution decomposition). Concern: violates repo Wire-Up Rule — new analysis modules need production caller or exp…

evidence: Verified: file does not exist on main. Test file has 11 tests. Wire-Up Rule cited from CLAUDE.md and AGENTS.md.

crypto_liquidity_wick_reversal_v1 has 366 rows split across 'battleground' (190) and 'Battleground_Main' (176) — same engine, two spellings. Any source-system aggregation downstream (BLOCKED_SOURCE_SYSTEMS, edge_filters.py, pf_registry source_share) will mis-count concentration. Fix: normalize at in…

evidence: SELECT source_system, COUNT(*) FROM at_raw_picks WHERE strategy='crypto_liquidity_wick_reversal_v1' GROUP BY source_system → battleground 190, Battleground_Main 176. Surfaced in FINDING_CRYPTO#1 inves

pf_registry.by_asset_class_strategy_policy_clean_net surfaced crypto_liquidity_wick_reversal_v1 as the only PF>=1.5 & n>=30 & WR>=50 strategy across all classes — looked like an edge. But all 30 'policy-clean' picks came from a single engine's closed_picks.json (battleground), with no cross-engine c…

evidence: Resolved via PR #86 (commit on branch feat/pf-registry-source-concentration). build_pf_registry.py now emits single_source_pct + top_source + is_single_source_artifact on every row in by_asset_class_*

Pipeline run 26641993453 hit OpenAI 429 across 3 retries on all personas for gpt4_1/gpt5_chat/gpt5_mini. Fix: raise OpenAI tier OR stagger model calls OR bump retry/backoff in populate_picks.

evidence: run 26641993453 log

config/model_persona_mapping.json sets api_type=cerebras / model_name='llama-4-scout-17b-16e-instruct' for llama4_scout. Cerebras returns 404 'Model does not exist or you do not have access'. Either pick a valid Cerebras model (llama-3.3-70b, llama3.1-8b) or route llama4_scout via OpenRouter (meta-l…

evidence: run 26641993453 Cerebras 404 response

Workflow env block wires ANTHROPIC_API_KEY: ${{ secrets.ANTHROPIC_API_KEY || '' }} but the secret is unset in repo, so the env var renders empty and populate_picks.py returns [] for cursor_agent silently. Fix: set ANTHROPIC_API_KEY in repo secrets.

evidence: run 26641993453 log shows ANTHROPIC_API_KEY: (empty) while OPENROUTER_API_KEY/OPENAI_API_KEY render as ***

tools/audit_pick_funnel/cli_track.py cmd_finding: when called to update an existing row by title (e.g. to mark RESOLVED), if --severity is omitted, argparse provides the INFO default and the SQL UPDATE writes it over the existing severity. Same issue likely affects --status. Fix: only include --seve…

evidence: tools/audit_pick_funnel/cli_track.py cmd_finding handler; reproduced 2026-05-29 22:18 UTC during the FINDING_OVERALL#1 update flow.

model_name 'z-ai/glm-4.7-flash:free' invalid on OpenRouter. Switch to a supported variant (e.g. z-ai/glm-4.6 or remove :free suffix).

evidence: run 26641993453 OpenRouter response

On 2026-06-02 the EAGLE-6 statistical admissibility gate was deployed to main via commit 2b4d7ce36 (code: alpha_engine/eagle_gates.py + 6f2.7 wire in alpha_engine/production_scanner.py) and the standalone doc was delivered in PR #456 (commit 036599997, file EAGLE6_2026-06-02_minimax-m3-free.MD). v1 …

evidence: alpha_engine/eagle_gates.py:165-356 (EAGLE-6 block); alpha_engine/production_scanner.py:5587-5599 (6f2.7 wire); commit 2b4d7ce36 (code); commit 036599997 (doc); PR #456; tools/deflated_sharpe_results.

active_picks_sync's LIMIT 5000 ORDER BY signal_timestamp ASC means with 39,706 OPEN CRYPTO rows the oldest-5000 window cuts off at ~Apr 7. The wick strategy is the canary, not the only victim — every newer CRYPTO strategy is similarly starved. PR #87 (max-rows 20000) is a partial fix; real fix is OR…

evidence: Subagent diagnosis 2026-05-30: 39,706 OPEN CRYPTO, 42 closed since Apr 29 (19W/23L), 2,403 legacy closed. See PR #87 + FINDING_CRYPTO#1 investigation.

at_raw_picks shows 366 picks for this strategy (190 from battleground + 176 from Battleground_Main), ALL status=OPEN. Meanwhile battleground/data/closed_picks.json has 43 resolved snapshots (25W/18L = 58.1% WR) — this is what pf_registry surfaces as the strategy's 'edge'. Either (a) TP/SL targets ar…

evidence: Surfaced in /money-maker-readyv2 audit 2026-05-30; reports/weekly_filter_2026-05-30T201255Z.md §3. Reproducer: SELECT status, COUNT(*) FROM at_raw_picks WHERE strategy='crypto_liquidity_wick_reversal_

Per top_edges_per_class.json (generated_at 2026-05-29T06:38Z — file is itself stale, see OVERALL #30): one CRYPTO segment passes BOTH Bonferroni adjustment AND walk-forward holdout: - Cell: trust=UNK & rr=RR1.0-1.5 & dir=LONG - n=327, wins=199, WR=60.86% (shrunk 60.23%), PF=3.885, avg_pnl=+1.09% - t…