Does the evidence support the next step?

One prespecified binary outcome, one independent randomized user per unit, two arms and a fixed final horizon. Absence of an event means no occurrence only when the analyst declares complete coverage. Continuous, ratio and clustered metrics are unsupported. No adjustment for peeking, specification search or interference. Funnels and exposure views are exploratory.

Blocking integrity problems, allocation mismatch (SRM p < .001) or an observed adverse increase above tolerance → Investigate. Missing declarations, immature windows, fewer than 50 mature users per arm, absent guardrail or uncertain estimates → Insufficient evidence. Otherwise, the primary lower 95% interval must clear the meaningful threshold, Fisher p must be < .05, and the guardrail upper two-sided 95% endpoint must stay within tolerance. All conditions are required. Fifty is an operational gate, not a power calculation.

Rates: Wilson 95%. Treatment-minus-control difference: Newcombe 95%, without continuity correction. Fisher exact: two-sided probability ordering on [[control outcomes, control non-outcomes], [treatment outcomes, treatment non-outcomes]]. Methods can disagree; review requires both. No corrections for multiple candidate outcomes or repeated looks.

SciPy Fisher reference · statsmodels interval reference

Imported content is processed in your browser. The app intentionally makes no data requests or application-storage writes. Cloudflare receives normal page-request metadata. Deliberate downloads persist on your device. Reset releases application references; it does not guarantee secure erasure from browser memory or disk.

The full replay pack contains raw input, mapping and specification fingerprints, methods, audit rules and results. Reopening recalculates from raw evidence and verifies fingerprints. Hashes detect changes, not authenticity. Use anonymized exports; inspect downloads before sharing. Reloading the page loses the current workspace.