R92 original/independent actual-return fragility source evidence¶

Failure-inclusive, frozen synthetic source study. Research-only record of two independent numerical implementations, not a production estimator or identified real-data return model.

Original 14,336-case archive¶

The complete archived original study contains cases.json, summary.json, source-commit.txt and SHA256SUMS. Eight separately recorded four-block original source shards are preserved. No missing, duplicated or selectively excluded source-method record is acceptable.

Separate independent actual-original source replay¶

R92-A independently regenerates every source and full missingness/secondary/assignment/actual-return mask with separately coded equations and frozen random streams; it does not import the original R92 source/estimator or the public package. It verifies X/Y arrays, every original SHA256 field, all 14,336 original source records and every fixed-cell signed bias, empirical error, oracle bias, count and bounded-mean width. The original archive's checkout and checksum members must match; fabricated, truncated or substituted source summaries fail closed. Exactly 32 independent complete generating blocks are preserved per fixed native source design.

The independent audit passing demonstrates reproducibility of the specified synthetic law and arithmetic only. It does not make the latent response probabilities observable in real eye-tracking data.

Frozen representative results¶

For matched original-source 50% invitations, the nominal uniform-completion adjusted HT oracle conditional design biases were:

The true uniform-completion source expectation for the nominally adjusted estimator has exact zero oracle conditional bias for all these nominal failure levels. Under outcome-selective failure the latent full-source oracle bias is nonzero, but the realized eight-draw empirical error need not agree in sign or magnitude due to randomization. These source-level values do not estimate a hidden completion propensity from observed data. A small observed RMSE does not validate uniform response.

For example, at 50% audit assignment and 5% failure, the original nominal-corrected finite-source empirical RMSE was approximately 0.01538 under selective versus 0.01543 under uniform completion for MAR independent curves, and approximately 0.01249 versus 0.01211 for MNAR repeated-trial participants. Similar RMSE cannot establish return-law equivalence.

Documentation publication and release boundaries¶

The R92 article, methods and two original-summary SVGs (analytical bias, empirical RMSE) are documentation-only. The website publication gate is a separate strict docs build, 25/25 exact-head PR workflows and successful protected-main GitHub Pages deployment. Never infer website publication merely from the original or independent scientific CI.

The source experiment does not identify field MNAR selection or return parameters, unknown raw missing gaze, true functional regression slopes, or 90% simultaneous functional confidence band coverage. Scientific blockers #296 and #289 remain open. Research source PR #453 and separate independent audit PR #454 remain draft, unmerged; stable 1.1.0 and public API unchanged; v1.2/GitHub Release/PyPI HOLD.