W3 R88 / R88-A — original and independent source evidence ledger¶

Scope: research-only conditional sensitivity of fitted selection coefficients under selectively returned randomized gaze audits. No released package changes, no general MNAR identification, and no 90%-band calibration claim.

Frozen original evidence¶

Eight four-block simulation shards and the three Python-version source contracts were mandatory predecessors of the original final SHA256 aggregate. Zero rows were dropped to produce the reported group means. R88 does not report cluster-valid uncertainty intervals or a formally estimated identified set.

Frozen independent numerical audit¶

The independent script imports neither original R87/R88 source estimators nor the production package. It recreates the frozen source X/Y, covariance and repeated-trial process, first-stage observation masks, secondary-recovery and third-stage randomization/return, and all recorded original 1,024 coefficients and refusals. It re-estimates the secondary selection model using SciPy score roots and the first-stage δ by a separate Newton solver. It recomputes original cell summaries and strictly verifies original SHA256 members and fixed checkout.

The first R88-A test attempt failed solely because an inherited error message said R87-A; the failure was retained and the label/fixture was corrected before the successful exact-head 4/4 replay. This is a transparently recorded engineering correction, not a scientific estimator change.

Original evidence contrasts (25% audit fraction)¶

Every displayed estimate is an unweighted arithmetic average of 32 source-level fitted δ values, not an individual gaze-record mean or formal coverage estimate. The four γ scenarios reweight the same observed returns; they do not represent four independent synthetic source cohorts. The true γ=1.35 is privileged simulation knowledge.

Figures and publication gate¶

The documentation-only website branch contains two source-derived SVG visualizations, R88 selection sensitivity and actual audit yield. The research article and methods companion preserve the design's scope.

The original research source PRs remain draft. Documentation publication requires its own strict exact-head docs and general checks, followed by a successful protected-main GitHub Pages deployment; source code/release remains excluded. Independently implemented numerical reconstruction does not establish historical gaze recovery, identifiable outcome-dependent return coefficients, 90% simultaneous-band coverage, or release readiness.

Keep #289 and #296 OPEN. Public 1.1.0 and API are unchanged. Version 1.2 / GitHub Release / PyPI = HOLD.