A3: absolute and relative grids (inspected)
Completed: 25 A3 folds and three prospective endpoints. Neither extension earns promotion. A3 remains the simpler reference; noise closure, convergence and sub-per-mille route stability are unaccepted.
All-five K5 grids and measured/model source browser · Independent spatial-start control · Complete method and chronological record · Independent confirmation audit
Same 2,249 fixed held cores, 50 dependence components. Both contrasts use the same 10,000 component draws; each two-sided 97.5% interval provides the declared 95% conditional family coverage. Positive means worse.
| Contrast | Median paired log score ratio | Conditional 97.5% interval |
|---|---|---|
| CCD19 / A3 | +0.00374233 | -0.00225252 to +0.00901412 |
| A3w / A3 | +0.01848306 | +0.00709430 to +0.03024686 |
CCD19 shows no demonstrated improvement. The frozen-width A3w recipe is conditionally worse. This is a comparison of the declared stopped recipes, not proof of converged-family equivalence or rejection of a physical width mechanism.
Equal-star mean residuals on radius-3 cores; all arms use the same per-star A3 fitted flux denominator. Catalogue T and recorded fitted colour split at the frozen training reference. Absolute rows share one scale, differences another. Common central and dipolar structures remain.
| Catalogue T | N | A3 | CCD19 | A3w |
|---|---|---|---|---|
| 8–9 | 52 | 1786.35 | 1799.18 | 1970.07 |
| 9–10 | 124 | 575.33 | 623.06 | 649.16 |
| 10–11 | 296 | 183.63 | 180.15 | 185.14 |
| 11–12 | 556 | 48.15 | 46.95 | 50.08 |
| 12–13 | 1221 | 15.17 | 15.22 | 15.65 |
Score is each source's mean squared supplied-error-standardized native residual, with no degrees-of-freedom correction. Absolute fits remain far from the supplied-noise scale. Fresh header/epoch/source-role population differs from development; these are not controlled cross-frame differences.
Evaluation amplitudes and global nuisance-prior scale are re-estimated on the evaluation image. Inference is conditional on those scales, fixed upstream preparation and learned models. Native quasi-gradient policies remain. A3/CCD19 hit the stage2 cap, then stage3 scalar plateau; width stopped on scalar plateaus. High losses and nonzero quasi-gradients prevent convergence claims.
Legacy fitted-flux T / BP-RP bins and own-arm normalization; use the matched image above for primary comparison.
All-five K5 audit verified 49,541 unique scene rows and one out-of-fold prediction per row. The in-training comparator is the median of the other four current-generation models on the same pixels; it changes training spatial coverage and initialization. One negative C4 amplitude remains in the primary cohort and has an explicit raw-stamp display exception. The alternate start tests spatial initialization while retaining empirical fine structure; it does not establish sub-0.5-per-mille route agreement.
Conclusion: retain A3 as the mandatory simpler reference. The completed bounded comparison does not accept noise-level accuracy or convergence.