Susceptibility-graded Release Control: Preventing Incomplete-cross-section Leakage in Financial Machine-learning Pipelines without a Blanket Timeliness Penalty

Many financial signals are same-period cross-sectional estimates-industry-relative ratios, abnormal accruals, any characteristic defined as the residual of a peer-group regression. Because firms file on staggered dates, a signal computed shortly after a fiscal period closes is estimated on the early-filing subsample; when filing latency correlates with the regression disturbance, the estimate embeds forward-looking selection ("incomplete-cross-section leakage"), a channel that standard temporal train/test hygiene (purging, embargoing, as-of joins) does not address. Existing remedies are all-or-nothing: release at announcement (biased for susceptible signals) or withhold every signal until a statutory deadline (a blanket timeliness penalty). We present a release-control system that grades the required cross-sectional completeness per signal from a directly estimable susceptibility primitive-the partial correlation between filing latency and the complete-cross-section disturbance given the construction regressors. Benign signals release early; susceptible signals are withheld until the cross-section is complete enough to suppress selection bias, up to the deadline. On a known-truth simulation-with the susceptibility primitive fitted on prior completed periods and frozen before gating (no oracle)-the controller releases benign signals at 36-39% completeness (vs. a 35% floor) with no detectable systematic bias, and escalates strongly susceptible signals to the deadline endpoint, inheriting its zero bias (naive-release bias -0.39). The companion analysis (Wu and Wu, 2026b) shows on observed U.S. SEC filing timestamps that the primitive separates signals whose apparent distortion is benign observable composition from those with irreducible selection; here that separation drives the gate. The controller ships as a reference implementation and is designed to sit upstream of feature-store materialization.