Evidence for small language slices¶

language_uncertainty_report accepts registered-language and controlled-label slices with local aggregate successes/trials for each declared stratum. It reports Bonferroni-adjusted Wilson intervals weighted by observed stratum proportions, not a pooled estimate that hides a tiny critical subgroup. Empty strata remain empty and force insufficient_evidence.

The minimum-detectable-effect summary uses a normal approximation for two equal, independent Bernoulli comparison arms, with conservative variance 0.25 per arm. Pin confidence, power, desired effect and minimum stratum size. Every stratum must meet both sample and effect-size conditions before ranking_allowed becomes true. Small perfect samples retain uncertainty and never imply language support. MDE is a planning approximation, not a retrospective claim of achieved power.

Confidence bounds are descriptive; sampling independence and representative strata must be justified outside this helper. Observed weights are not population weights. No model runs or network calls occur. Outputs contain only controlled language/label/stratum codes, aggregate counts and statistical summaries; codes must come from an approved vocabulary, not source identifiers. The helper cannot certify privacy, clinical performance or a model release.