Post-Commitment Leakage Probe

(1) Integration point: use this as an evaluation wrapper around a temporal model, not as a gradient update. For each example or interaction, define a hard commitment time. Record the model input tensor and prediction immediately before commitment, then request a later exogenous delay, scheduling offset, sensor latency, or randomized presentation interval. The endpoint target must be fixed by the commitment event; no post-commitment tensor, timestamp, mask, batch statistic, or label-derived preprocessing may enter the prediction. Freeze the model, tokenizer, normalization statistics, fold assignments, and retention rule before reading the randomized delays. (2) Pseudocode: split data into randomization strata; fit the model on training folds only; for each held-out item j compute pred[j] = model_frozen(X_pre[j]), residual e[j] = A[j] - pred[j], and retain item j using a predeclared rule. Compute T_obs = sum((tau[j]-mean_tau)*(e[j]-mean_e)). For b in 1..B, permute tau independently within each stratum and recompute T; return the two-sided permutation p-value from the displayed formula. A signed one-sided test can target the direction predicted by the alternative. (3) Mathematics versus estimation: the conditional-independence null and retained-sample requirement come from the paper; residuals, stratum means, and permutation quantiles are empirical. Check positivity by requiring every retained stratum to contain at least two delay values and report strata with poor support. (4) First experiment: construct a synthetic sequence benchmark in which the target is generated from tokens before time t, while a random delay is assigned after t. Compare a leakage-free Transformer with versions receiving a future token, absolute final-sequence length, or batch-normalization statistics contaminated by future examples. Use 10k sequences, a small 2-layer Transformer, and 1,000 within-stratum permutations. The expected signal is a calibrated null p-value for the clean model and a reproducibly small p-value plus nonzero signed T for each leakage variant, while ordinary validation loss may remain almost unchanged.