W3 R89/R89-A — independently replayed finite-population bound evidence¶

Frozen source and independent numerical provenance ledger. This is not a public estimator or an inferential calibration claim.

Original source experiment¶

The original workflow required passing Python 3.11, 3.12 and 3.13 source/contract tests, eight independently generated four-block shards and a failure-inclusive checked aggregation. All 512 prescribed source-method records, counts and scientific restriction flags are retained.

Separate independent verification¶

The independent audit script does not import the original R89 generator, missingness mechanism implementation or package. It regenerates X/Y and the original random streams, observation masks, secondary returns, randomized assignments and outcome-selective audit completions. It recomputes the sharp observed-only endpoints, contrast endpoints, source-level original truths and aggregation.

Failure-inclusive audit disclosure. The first independent replay correctly refused an incorrect expected checkout SHA because the test had pinned the source branch head rather than the actual GitHub pull-request merge checkout. The failing CI was retained. After correcting only the provenance pin, test fixture and documentation, the full mandatory actual-archive numerical replay passed. The correction made no change to original source data, estimator assumptions or scientific results.

Original point summaries¶

The following are source means of sharp interval widths for the bounded outcome Z=clip((Y+1)/2,0,1), not mean estimation errors or nominal confidence coverage.

Corresponding predictor-group contrast interval widths range from approximately 0.703 under 50% complete audit return in the participant MNAR cell to 1.200 under 25% selective return in the curve MAR cell. 0 of 32 independent generating blocks in every fixed cell produced a contrast interval excluding zero.

The true source transformed-outcome mean and predictor-group contrast were contained in every original interval by finite-population worst-case algebra. This is a deterministic source integrity test—not frequentist confidence coverage. The independent re-estimation verifies agreement with the exact frozen source conditions; it does not prove MNAR point identification.

Figures, website and release boundaries¶

The accompanying R89 article, methods companion, bounded-mean width SVG and contrast-width SVG are a documentation-only scientific publication. Publication requires its own exact-head strict docs CI and protected-main Pages deploy; original source PR #444 and independent audit PR #445 remain draft and unmerged.

No raw-gaze or arbitrary function-on-scalar slope identified set, no missing-gaze reconstruction, no latent audit-return model identification and no simultaneous 90% band coverage claim are authorized. Issues #289 and #296 remain OPEN; stable 1.1.0 unchanged and v1.2 GitHub/PyPI release HOLD.