W3 R86: original and independent third-stage failure audit evidence ledger¶

Research-only numerically source-grounded evidence · 21 October 2026

Original immutable source and independent audit¶

R86 sources use entirely new RNG seed streams: complete source 202610218680, original MAR/MNAR masks 202610218681, secondary validator 202610218682, third-stage independently randomized audit 202610218683. The original science and independent audit independently reconstruct all source X/Y, mask, recovery/audit uniform and pilot SHA256 provenance. The first implementation uses Newton recovery estimation and bounded-bisection first-stage delta; the independent implementation uses SciPy multidimensional score-root recovery estimation and scalar Newton first-stage solving.

Figure 1 — two-stage selection and the original MAR/MNAR controls¶

Figure 1. Mean first-stage fitted original missingness selection delta in the two-signed-extreme predictor heteroskedastic source, n=32 new independent blocks per fixed cell.

Values are source means, not inferential bounds. These are counts of time×coordinate×trial recovered measurements, not additional independent participant sample sizes.

Figure 2 — estimating the second-stage recovery selection coefficient¶

Figure 2. True source gamma=1, 2-anchor heteroskedastic source, independently reproduced original source means.

At zero audit the audit-estimated method refuses recovery coefficient estimation, so no source mean is presented. At positive q, source means differ from truth due to finite samples and pilot-model dependence. A weighted score solution is not evidence of a valid sampling covariance or interval.

Figure 3 — all eight source cells, including unfavorable performance¶

Figure 3. First-stage original MNAR generating truth delta=1.1. Each value is a mean over 32 new independent source blocks; no cell excluded.

More randomized auditing adds genuinely measured outcomes, yet source-mean absolute error does not improve monotonically in every cell. In the adverse two-anchor heteroskedastic source, independent curves show .102905 at q=.25 versus .114480 at q=.50; participants show .125260 versus .128455. These are not confidence-interval coverage comparisons. The zero-audit naïve control also has a different weighting assumption from audit-fitted positive-q strategies.

Scientific release decision¶

R86 independently reproduces the specified synthetic data-generating law and the information effect of a third-stage randomized audit among otherwise missing samples. It does not justify a claim that the secondary outcome-selection model, its measured failure set, or a historical lost eye-tracking response can be observed or recovered in practice. It does not identify a general MNAR law, validate first-stage functional slope inference, or resolve the original nominal-90% coverage blocker.

Protected main estimator/API unchanged; stable eyetrajectoriespy 1.1.0 unchanged, scientific issues #289/#296 OPEN, eyetrajectoriespy 1.2/PyPI RELEASE HOLD.

Full R86 article · Methods companion · Prior R85 secondary-selection falsification