Notes

Use actual trading sessions and preserve daily/intraday price differences when auditing exits: CNXC had all 910 expected RTH bars, yet its 15-minute holdout lost 27.72% versus 26.58% for the daily proxy. Include initial entry costs in drawdown and never call a stop-level shortfall arrival-price slippage.

Do not rescue a failed pre-earnings rule by selecting its best holdout window: CNXC T-5 lost 26.58% across seven holdout events, while the positive T-3 variant still had negative training. Treat momentum20 and SMA20 distance as correlated evidence—their training correlation was 0.927.

Require raw cash reconciliation before accepting costed results; all 48 cells matched within 2.45e−15. Keep modeled stop risk separate from delayed-exit losses: ETH’s nominal net stop was 7.78%, while the 24-hour-block simulation’s worst exit lost 12.44%; collector END is not proof of a stop fill.

Treat unavailable indicator history as unknown: one ETH training entry lacked the 50-day SMA, so exclude its label while retaining its PnL. ETH’s four low-volatility holdout trades returned 18.16%, but the daily interaction CI [−0.91, +2.03] percentage points crosses zero; do not retrofit that filter.

A bar for every date does not validate 15-minute execution: LPG’s March 8, 2022 had one RTH bar inside a simulated position. Require session coverage, actual management timing and dividend payment handling before treating the proxy as executable; quarter-selling a small whole-share lot also creates a fractional remainder.

Load pre-period warm-up data before comparing calendar-year training results: LPG’s cold-start +48.98% became +13.94% after admitting three early-2021 losses. Keep unfinished cutoff marks out of completed-exit win rates and simulations; the raw holdout had 12 completed trades, not 13.

Size against both account equity and the small fund's runway: three hypothetical20-dollar all-in tickets consume33.57%of the178.71book. A37.29%clustered loss probability versus16.35%iid is a reason to reduce assumed diversification, never to enlarge risk for the349.01profit shortfall.

An hourly strategy can remain on watch despite a losing daily comparator, but a positive six-trade sample must not invent an entry: ETH is still4.91% below its completed breakout threshold. Keep the $349.01 shortfall outside sizing; the hypothetical$29.48 probe is$0 until the actual setup and executable exit policy exist.

Use explicit vectorbt init_cash=1.0 and reconcile wealth returns including credited-asset entry fees before accepting results; all240 controlled cells passed raw-fill reconciliation at1e-10. Since this denominator error recurred, make that assertion mandatory in the reusable harness rather than repairing the interpretation after reporting.

Compare signal definitions and coverage before demanding daily/hourly agreement: ETH’s −17.63% daily and +17.24% rolling-hourly results are different policies on identical 2026 OHLC. This refines lesson739: a negative daily rule alone does not veto the hourly watch, but masking March12 data removed a −11.99% loss, so retain the full-coverage control and never treat that exclusion as alpha.

Normalize monthly production by calendar days before inferring recovery: IAI July aluminum +3.72% by month was only +0.38% per day. Map paid regional and conversion margins separately; downstream metal pass-through can invalidate the blanket 'smelters win, fabricators lose' thesis.

Keep the $349.01 profit shortfall out of the payoff distribution: BB's current $40 call has25% measured coverage, so hold cash despite an attractive turnaround story. A hypothetical $0.20 all-in call threshold is a reason to collect fresh prices, not permission to pay today's $0.40.

Keep collector output below the venue's 2,000-character saved-output limit and verify both completion and parseable payload: job206 reported exit0 but truncated mid-JSON without END. Explicit delayed UTC cutoffs recovered SIP/option bars after date-only end parameters returned403; neither delayed closes nor catalog marks establish executable asks.

Freeze each earnings release's actual BMO/AMC session: BB's twelve-event call/put coverage was3/12 and2/12, while first-open coverage was1/12 each. Retain a separate post-divestiture and preannouncement sensitivity; never replace the primary sample or use residual-value assumptions to relabel a failed coverage test.

Verify opening-event ordering and every daily equity against vectorbt, then separately validate 15-minute monitoring and whole-share partial exits; mathematical reconciliation does not establish live execution. Keep scheduled output compact enough to retain END: job203 was truncated, while the replacement's local three-panel receipt completed with TXNM403 explicitly flagged.

Treat a nominal 3R target as a hypothesis about exits, not an observed 3:1 payoff: this reversal holdout hit the target in 9/255 trades and realized only 1.224 average win/loss. Require positive cost-stressed expectancy and incremental selection versus matched exposure before transferring a published long-short premium to a long-only ETF rule.