W3 R84: independent actual source/ascertainment evidence and original figure ledger¶

Scientific research evidence · 21 October 2026 · synthetic independent recovery of otherwise missing Y only

Original and independent immutable evidence¶

The new source RNG streams were prospectively fixed at 202610218480 (complete X/Y source), 202610218481 (original MAR/MNAR masks), and 202610218482 (independently randomized external recovery). The original source family retains correlated Gaussian functional trial innovations, independent-unit Student-t5 radial shocks and rare signed asymmetric unit shock.

The visible-only mean pilot never uses newly rescued outcomes or generating truth. The scalar selection-model likelihood receives Y only for originally visible or independently recovered missing samples. The true generating δ is used only for source performance evaluation, not model fitting.

Figure 1 — identified-information contrast and source controls¶

Figure 1. Source-mean fitted parameter under the primary two-extreme independent-unit, predictor-heteroskedastic condition. Dashed true-source δ values are for simulation validation only.

All positive-recovery cells above used 32 independently generated sources and had 32 converged scalar coefficient fits. A converged point estimate is not proof that the model was correctly specified, nor a confidence interval.

Figure 2 — all eight prespecified MNAR source condition errors¶

Figure 2. Absolute fitted δ minus true simulator MNAR δ=+1.1, averaged over 32 genuinely new independent generating blocks in each cell.

Higher recovery often reduces source-mean realized absolute error but is not uniformly beneficial in every small finite-source cell. That is compatible with noisy recovery, source-level dependence and the visible-only pilot being an imperfect conditional-mean specification; the experiment has no interval or consistency guarantee.

Figure 3 — independently measured formerly unobserved Y¶

Figure 3. New observed coordinate/time/trial values supplied to a synthetic selection estimator, not additional independent people.

For each original complete source, the 25%-recovered otherwise missing observations are a subset of the 50%-recovered observations through one independent recovery uniform field. 0% supplies no new outcomes, so neither the MAR nor MNAR δ is fitted at all.

Scientific conclusions that remain unsupported¶

The archive reproducibly answers whether a particular parametric selection score, using a visible-only pilot and additional independent synthetic missing-response ascertainment, yields informative scalar fitted δ values in these source conditions. It does not show that lost gaze can be externally remeasured, that random ascertainment can be implemented on a real participant's historical gaze, or that the recovered-outcome model remains valid with covariate-dependent validation, misclassified samples, unknown recovery fractions or a misspecified MNAR family.

The conditional scalar coefficient is not the complete time×coordinate functional slope and it supplies no full-function inference covariance. No nominal 90% calibration, source-independent interval coverage, or nonparametric MNAR identification is established, even though all original synthetic source diagnostics have been independently recreated.

Read the R84 scientific article · Methods companion · Previous R83 five-assumption sensitivity experiment