Red-team checklist

• Reconstruct each signal timestamp and prove all inputs were available at formation close.
• Verify thresholds use prior observations only, forward labels start at t+1 and complete horizons stay inside chronological splits with a 28-day embargo.
• Audit point-in-time ranks, listing/delisting history, migrations, renamed tickers, missing index constituents, and whether missingness selects survivors.
• Quantify overlapping windows, event counts, effective time blocks and common shocks; use block uncertainty.
• Report base/signal event rates, risk differences, risk ratios, confidence intervals, MAE severity, worst-5% ES, asset counts and all exclusions.
• Apply Holm across all 2,592 exploratory grid endpoint tests. Apply Holm to the two anchor co-primary endpoints inside each validation, historical holdout and sealed forward block.
• Check placebo signals against already-realized crashes, cluster-preserving permutations, and that thresholds exclude current/future observations.
• Compare the registered model to predeclared prior-only controls; never choose the model with the best p-value.
• Test 28/56/84-day bootstrap blocks, regimes, liquidity cohorts and other free venues as robustness, clearly labelled exploratory unless preregistered.
• If effective samples or history are too small, conclude inconclusive. Do not relax the gate after seeing outcomes.
• Predictive association is not causality and is not evidence that a trade or short position is profitable.