W3 R90/R90-A — original-source audit saturation evidence ledger¶

Frozen, independently replayed synthetic experiment. The data compare assigned versus actually completed audits; all inferences are bounded finite-population descriptions, not nominal confidence procedures.

Immutable original study¶

The original archive has cases.json, summary.json, source-commit.txt, and SHA256SUMS. Eight separately uploaded four-block source shards are retained. All records are required; aggregation fails closed for source duplication, missing cases, mixed checkouts or modified members.

Independent original-archive replay¶

The R90-A script does not import the original R90 source generator, bounds implementation or public package. It independently reconstructs all original X/Y arrays, first-stage masks, validation, nested third-stage audit assignments, both return laws, complete and measured outcome digests and finite-population sharp bound endpoints. It verifies all original 1,280 source-method rows, original source checkout and SHA256 members, independent-unit denominators, strict width monotonicity, zero/full-assignment boundary conditions and first assignment fractions meeting widths 0.10/0.05.

Failure-inclusive record: The initial independent actual-archive replay did not reach numerical verification because its audit checker referenced an inherited R89 summary-field name. That failed run is preserved. A traceable audit-only mapping fix was committed and the three-version fixtures and mandatory actual original full-archive replay reran successfully. No source selection, estimator, random seed or original archive data was modified.

Independent-source descriptive summaries¶

The mean interval widths for q=100%, averaged over 32 independent source blocks within each fixed cell, are:

Complete return reaches both thresholds in 32/32 source blocks per fixed cell; selective return reaches neither threshold in 0/32. Original q=0 widths range from about 0.699 to 0.729 by first-stage law and native design, and are identical between policies at q=0.

For a bounded-transformed predictor-group contrast among originally missing records, the q=100% ideal complete-return interval collapses to a point, with the realized contrast nonzero for the archived 32 blocks per cell; under outcome-selective completion, 0/32 blocks per fixed cell achieve a contrast interval excluding zero. This is finite-population information, not inference about regression coefficients.

The mean width itself is exactly the unreturned fraction of an originally missing target population. It is not a nominal-90% interval, and its true-source inclusion is guaranteed by the [0,1] range, not empirically established sampling coverage.

Documentation sources and scientific boundaries¶

The R90 article, methods companion, saturation curve, and full-assignment threshold comparison contain only descriptive results from the frozen source summary and independently verified original archive. Their publication is documentation-only, gated by strict docs build, 25/25 pull-request CI and a separate GitHub Pages deployment after squash merge.

Research source and independent audit remain isolated in draft PRs #447/#448, not merged into protected main. The work does not identify MNAR hidden-outcome return coefficients, raw missing gaze, arbitrary functional regression slopes or 90% simultaneous functional bands. Issues #289 and #296 remain OPEN. Stable public 1.1.0 unchanged; v1.2 release/GitHub/PyPI HOLD.