SDR-0013: Report event selection before and after preprocessing

• Status: Accepted
• Date: 2026-07-26

Context

The outcome-blind IBL event-coverage audit found that the prospective regularization policy retained every frozen event, but the older rolling-validity gate had already removed 1,612 otherwise boundary-eligible events. Aggregate retention was 99.29%, while one session retained 88.1% and a different session had an 8.44-percentage-point condition-retention difference. Reporting only the final denominator would conceal where selection occurred.

Decision

Event-analysis workflows report at least three denominators when applicable: candidate events with nominal recording-boundary coverage, events retained by eligibility/QC gates, and events complete after preprocessing. Counts and retention fractions are stratified by condition, session, and animal. Upstream selection and incremental preprocessing attrition remain separate; a later stage cannot receive credit for events removed earlier.

These diagnostics are structural and should be available before fluorescence outcomes or effect estimates are accessed. Aggregate balance never substitutes for session- and animal-level reporting.

Consequences

Reports become slightly larger, but scientists can identify differential selection, distinguish acquisition/QC limitations from algorithmic loss, and reproduce the exact analysis population. Product APIs need a first-class coverage audit rather than a lone warning or final event count.

Revisit trigger

Revisit the required strata or default warning thresholds after moderated usability studies and external datasets reveal whether the report is too noisy or misses consequential selection patterns.