W3 simultaneous-band calibration: audited evidence and original plots¶

Evidence gallery · updated 20 October 2026 · synthetic Gaussian simulations only

Figure 1 — R71/R73/R75: estimated procedures compared with generating-law references¶

Read the scientific article: when numerically correct simultaneous bands under-cover.

This is a cross-study descriptive display, not a paired estimator comparison across R71, R73 and R75. Each reported cell has 32 independent generating source blocks; different stages generated different synthetic datasets, covariance regimes and/or estimated procedures.

All figure bars use truth-inclusion counts out of 32, not bootstrap draws or percentages silently substituted for counts. The generating Gaussian covariance is known only to the simulation. R71's extreme curve comparison, for example, is 2 vs 26 successes on different candidate/oracle methods applied to the same 32 original datasets, whereas R71 and R73 did not reuse source datasets.

Figure 2 — R76: whole-function truth inclusion by replicated extreme support¶

Read the scientific article: how independent leverage support affects coverage.

The independent predictor extremes x=±4 are 2, 4, 8 and 16 separately simulated independent units, evenly balanced by sign; other predictors lie in Uniform(−0.95,0.95). The adverse covariance regime scales unit SD by \(1+0.75|x_i|\). This is 32 independent generating source blocks for every design/support/variance condition; the 512 complete cases across all conditions share paired source blocks.

Figure 3 — R76: why source-wise SE diagnostics matter¶

The first panel plots the generating-law variance concentration index \(N_{\mathrm{eff,var}}=1/\sum_i w_i^2\), where \(w_i\propto a_i^2s_i^2\). The second plots the mean across 32 source datasets of each dataset's 10th percentile across 62 slope coordinates of \(\widehat{\mathrm{SE}}_{HC3}/\mathrm{SE}_{true}\). Both require simulation-known variance or SE, and neither is a validated plug-in diagnostic for real gaze data.

Increasing the number of independently contributing high-leverage units cochanges regression support, slope precision and source information weights. The observed improvements support a finite-sample information-concentration mechanism in this experimental law, but do not isolate a pure causal effect or define a universal minimum effective sample size.

R76 reproducibility and audit ledger¶

The plot artwork is a vector re-expression of the independently verified numerical evidence displayed in the tables above, not a new experiment, an independent replication or a plot of raw human eye-tracking recordings. The SVG files are checked into the documentation tree with the source-count ledger; any changes to the plotted scientific counts require checking against the separately retained source artifacts and independent audit report.

Reading order: Why simultaneous bands under-cover → R76 high-leverage replication → Scientific evidence dashboard. Original methodology remains in the research PRs; website publication is not research-code integration or product release.