Last-green multilingual regression gate¶

cross_lingual_regression_gate compares every expected language/label baseline with a proposed checkpoint using aggregate paired outcomes from the same pinned cohort. Supply baseline/candidate checkpoint digests and an explicit expected cell-to-cohort map. Missing cells, unexpected cells, mismatched cohorts and small samples fail qualification rather than disappearing into aggregate F1.

Each PairedTransferCell carries success totals and baseline-only/candidate-only success counts. The constructor checks their joint-count consistency. Statistical classification uses an exact two-sided paired discordance binomial test with Bonferroni correction. Exact summation is bounded to 4,096 discordant pairs; larger unbalanced cases report an explicit computation-budget limitation and remain statistically inconclusive. Balanced pairs have exact p-value 1 without summation. Critical decline policy still applies. Positive, negative and inconclusive cells stay separate.

An independently baseline-policy-declared critical cell whose point recall decline exceeds the pinned tolerance fails release independently of statistical significance. Target gains cannot offset it. This conservative policy does not call an inconclusive decline statistically established. require_cross_lingual_release enforces the result by raising a controlled error when qualification fails. Call it from a checkpoint promotion boundary; existing release jobs are not silently rewired by this module.

Reports contain controlled codes, counts, deltas, digests and statistical values, never source records or identifiers. Labels and critical designations must come from an approved local policy, and baseline/cohort provenance must be verified outside this API. Digests do not prove that a baseline is last-green. The paired sampling assumptions and multiplicity family must match the actual evaluation. No model execution, clinical certification or publication occurs.

The required baseline map binds every (language, label) to (cohort_digest, critical) from trusted last-green policy. Candidate critical flags must match; a downgraded flag fails qualification. The baseline map is bounded to 1–10,000 cells before iteration.