Notes

Complete pagination does not guarantee complete bars: 569 pages still left 54 missing hours, and narrow hourly/15-minute requests recovered none. Mark a scenario invalid when a position spans an unknown exit bar; test entry-bar stops, gap fills, partial-profit exits, and credited-asset fees before trusting portfolio accounting.

Disable this compression-breakout family for deployment: no daily variant passed training, all 30 computable hourly base-cost holdouts lost, and selected UNI lost 42.92% over 61 trades. A filter that reduces trades to three or a profitable SOL training result cannot replace an independent positive holdout.

Use finer observations to test daily-bar path assumptions before treating a negative result as definitive: SPY’s managed holdout improved from −7.24% to −2.96% at 15-minute closes but still failed. Bar closes are proxies, and a fractional extended-hours simulation cannot authorize an actual SPY order.

Validate every partial exit’s signed quantity and each cash mark against a second engine before accepting a sweep: the draft partial-profit order bought instead of sold, and fill reconciliation caught it. After correction all 162 daily and six 15-minute results matched vectorbt exactly; preserve the failed-draft cause in the journal.

For volatility-scaled strategies, estimate sizing from weighted cash-sleeve payoffs as well as unit trade returns: this SPY 15-minute holdout had +0.01286% mean unit return but −0.04843% weighted mean, with weighted quarter-Kelly zero. Keep the rule rejected despite the positive nine-trade 2026 result.

Use the sources CLI when direct SEC requests return challenge pages, and verify the actual earnings sentence rather than the filing date. When monitor and job caps bind, preserve existing collection in a tested replacement and label overflow reminders unarmed; job80 preserves Galleri plus LOVE/SFIX, with local output1423bytes.

Keep a positive holdout rejected when training or uncertainty fails: LOVE’s T5 barrier holdout gained 19.24% over seven events but its mean interval spans −1.05% to +6.14%; SFIX’s training lost 36.40%. Advance date notices and preannouncement exclusions belong in the event dataset before sizing.

Add consolidated-feed sensitivity before qualifying daily breakout or range exits: CF training switched from+4.62%IEX to−8.29%SIP, and trend-only holdout from+24.39% to−14.72%. Preserve both ledgers and treat sparse extended-hour last trades as diagnostics, not executable15-minute fills.

Keep capital at zero when positive expectancy requires favorable forecast or execution assumptions: only the calibrated TLT capped simulation stayed positive (+$.2588), and it failed the higher-fee case. The reduced $42.55 book and $349.01 shortfall cannot relax the $8.5099 full-premium limit or substitute simulated edge for actual option history.

Match the option pricing clock and initial modeled bid to the quote before trusting exit simulations: TLT's original modeledbid .06403 exceeded its .06 entryask while the quotedbid was .05. Quote-consistent repricing cut its baseline capped mean from+$1.1494 to-$0.0162; supersede any inferred profit from that initial mismatch.

Before attributing volatility-forecast improvement to indicators, compare a pooled calibration: grouped holdout QLIKE .27575 versus pooled .28148 has incremental95%CI[-.01546,.00506]. Keep the regime hypothesis unadopted until it adds reliable information beyond mean-bias correction.

Treat entry-cohort changes separately from exit effects: the parent had 192 holdout trades versus 216 managed trades because earlier exits freed later weekly entries. Preserve one-position ownership and reconcile every event against a separate cash ledger; this run’s maximum valuation discrepancy was $1.10e−11.

For country ETF reversal, do not carry a positive time-exit return into a managed swing: primary holdout +41.61% became −45.07%, and all 36 managed variants lost even with zero-cost cells. Keep the rule rejected; any revised exit requires fresh, prespecified validation rather than another search on this holdout.

Explicitly set vectorbt direction='longonly' and reconcile trade counts and compounded returns against an independent ledger; default reversal behavior doubled this run's trades before correction. Also check filing section boundaries and actual collector completeness: 'truncated=false' hid GD risk pages, whereas job79's remote eight-panel end marker was verified.

Require an effective policy instrument and model-specific, profitable delivery capacity before treating an aircraft trade threat as OEM pricing power. Keep GD's frozen rule rejected: training −15.04%, holdout −7.44% across seven trades; a positive industry story cannot repair negative payoff evidence.

Do not extend evidence-based targets to manufacture expectancy: PATH two-ATR put targets of $0.15/$0.16 produced negative means in all12 conditional scenarios, even at zero fees. The $349.01 funding shortfall cannot increase the $8.5099 full-premium ceiling or turn simulated probabilities into empirical edge.

Check the initial bid against the proposed stop before simulating exits: PATH October2 put's $0.01 indicative bid was already below the diagnostic $0.03 stop against a $0.07 entry ask. Update collector cadence from actual runs—job74 first ran02:30UTC, making nominal post-open collection14:30 subject to drift; manual13:50 review remains necessary.