# Conditional colour-off control on new-full F1

Completed 2026-10-04 by model_campaign at the parent's request. This is a positive-control ablation of the completed strict-fold A3 checkpoint, not a separately fitted achromatic model, a third selection family or an independent confirmation exposure.

The learned spatial ePSF and WCS remain fixed. The off arm zeros the original17 colour coefficients and the already-zero extra CCD translation leaf. Both evaluations follow the existing exact adapter/scorer: source amplitudes are solved, the global nuisance-prior scale is recalibrated on the evaluation image, then amplitudes are solved again. The2,062 source IDs and their held cores are fixed from the parent paired contract; each has at least15 valid pixels within3 pixels of its original scene position. No arm-specific deletion occurs. ON scores and amplitudes exactly reproduce the stored A3 evaluation.

The median paired ratio of OFF/ON core mean-squared noise-normalized residual is **1.47877**. The conditional95% interval from10,000 resamples of the49 pre-existing dependence components is approximately **1.370–1.605**. This interval does not include uncertainty in the global evaluation calibration or noise model. The separate median core scores are25.8032 ON and44.2103 OFF; their ratio is not the median paired ratio. Input/checkpoint/adapter hashes and exact scalar values are in summary.json; paired source-level scores are in paired_scores.npz.

The primary figure shows the actual observed data, ON/OFF full scene model, residuals and OFF-minus-ON residual, using the same ON fitted amplitude as denominator and the same held-core support. Colour is the fitted Gaia-XP coordinate, split at the training reference0.955536; magnitude is catalogue Tmag. Legacy standard grids instead use BP−RP and ON fitted-amplitude magnitude. Their automatic colour scales differ, so use the shared-scale primary figure for direct visual comparison. All five figures were inspected.

![Observed images, models and held-core residuals](held_stacks.png)

Conclusion: removing the learned colour terms produces substantially larger held-core prediction errors and strong opposite-colour dipoles on this conditional evaluation. The existing colour model captures an important component of the data. Residual structure remains with colour terms enabled; this control does not establish noise-level closure, absolute flux accuracy, an optimal achromatic comparison, or a benefit from the additional two-coefficient extension.

Reproduction: activate syndiff, then run colour_off_control.py from the isolated codex/epsf-model-closure-20261004 worktree. Runtime is the immutable production3609d46 archive with adapter SHA49ad6a444f2470f203ec3526921c4f6cc861d11caabdb95214b95f16d0a267cb. No training was run or modified.

Row-identity audit: the full F1 scene has two repeated numeric IDs among nuisance rows. Neither appears in the fixed2,062-source held cohort, whose IDs are unique. The source order, original row selection, coordinates and ON scores match exactly, so the reported ablation is unaffected. row_identity.npz now records row ordinal, source ID, bundle index and input-scene hash. The reproduction script selects directly from the saved evaluation contract rows instead of a source-ID dictionary. Full metadata and the equivalent C4 check are in ../row_identity_audit.json. The origin of the repeated numeric IDs is not inferred from their proximity.