W3 R85: secondary validation selection, full original-source evidence and figures¶

Source and independent audit provenance · 21 October 2026 · synthetic research, not public inference

Original and independent numerical provenance¶

R85-A reconstructs all original real archived source-level results, rather than independently verifying only test fixtures. It uses centered-sufficient-statistic pilot OLS and independent Newton roots to compare against original trial-level least-squares and bounded bisection. Original X/Y, original masks, recovered masks, propensity arrays, pilot fits, final deltas, source-level error records and immutable hashes are independently verified for the entire factorial.

Figure 1 — primary adverse MNAR source δ¶

Figure 1. Two signed influential predictor units, predictor-dependent source variance; 32 independent blocks/design, true first-stage MNAR δ=1.1.

The MAR/MNAR first-stage generating truth is not used by fitting; the simulator-only recovery-propensity oracle uses true conditional mean only to calculate correct validation weights. These are descriptive mean point estimates, not inferential bounds.

Figure 2 — full eight-cell latent-selective source error¶

Figure 2. Mean source \(|\widehat\delta-1.1|\) across 32 primitive independent source blocks/cell for actual original first-stage MNAR masks and the same recovered samples under both weighting strategies.

The oracle's lower source-mean absolute errors in these eight cells are a synthetic diagnostic, not evidence that the oracle recovery propensity can be identified from actual observed gaze.

Figure 3 — first-stage MAR, validation-stage MNAR produces a false coefficient¶

Figure 3. First-stage generator true δ=0, 2-anchor heteroskedastic source, 32 new independent blocks/cell.

The source with MAR original gaze selection does not become truly MNAR merely because a secondary validator preferentially recovers certain hidden residual values. Ignoring that second selection can nevertheless lead to a fitted positive original MNAR coefficient. These are not formal significance tests or false-positive probabilities.

Qualifications of secondary observed-input weighting¶

The recorded-selective corrected strategy assumes the validation law is known, logged and governed by recorded predictor/time. It is not generally guaranteed to reduce finite-sample mean absolute first-stage δ error relative to an incorrectly constant recovery weight. For the adverse 2-anchor heteroskedastic original MNAR participants, source-mean absolute δ error was .1189 under naïve versus .1390 under stipulated correct recorded-input recovery weights, although source-mean signed fitted coefficients shifted from 1.1862 towards .9610 (truth 1.1). Error magnitudes and signed shifts address different criteria.

No source-independent uncertainty, estimated participant-cluster covariance, valid missingness-adjusted functional coefficient or calibrated whole-trajectory confidence band is computed here. The validation mechanism is itself a scientific object requiring independent information and measurement justification.

The full R85 article and methods companion explain the falsification and its limits. The original archived datasets, individual original source/mask/recovery hashes and all 2,560 source-strategy diagnostics remain available in the linked numerical evidence, not extrapolated to real eye-tracker behavior.

Research source code stays draft, blockers #289/#296 OPEN, protected-main inference unchanged, version 1.2/PyPI RELEASE HOLD.