# Final ePSF magnitude/colour closure — 2026-10-04

**Status: completed bounded investigation, with scientific limitations. Manuscript writing remains paused.**

**Read the [final conclusion and evidence](FINAL_CONCLUSION.md).**

**Current conclusion:** current-template A3 retains structured held-out residuals. Neither the two added CCD colour shifts nor the historically fixed magnitude-width response demonstrates an improvement on the F1 development test. The reserved-exposure comparison also supports no extension promotion: the fixed width worsens its conditional score. Retain A3 for the tested recipe; no noise-limit or absolute-photometry acceptance is claimed.

[Evidence canvas](http://localhost:18888/epsf-model-closure/) · [Live work status](status.json) · [Remaining independent checks](review/FINAL_CHECKLIST.md)

## Completed evidence

| Question | Evidence and current interpretation |
|---|---|
| Are the latest templates actually used? | All five development scenes rebuilt from the latest native templates, method-A background and latest F4 Hotpants. Frozen source/pixel support is checked. [Producer records](models/inputs/) |
| What model is fitted? | Full spatial A3 with17 colour terms; CCD19 adds2 learned shifts; fixed-width A3w adds no learned nonlinear coefficient. [Exact equations, units and counts](models/FINAL_MODEL_CONTRACT.md) |
| Do the extra terms help? | F1 paired development intervals include no change for both additions; C4 CCD19 slightly worsens the conditional score. [Current F1 comparison](significance/current_F1/) and [C4](crossfit/C4_full_generation/fold0/paired/summary.json). |
| Are residuals consistent with pixel noise? | All three current F1 models fail the declared conditional fixed-profile reference. Shared pixels, flux refitting and multiple checks are accounted for; learned-model/preparation uncertainty is not. [Images and full scope](significance/current_F1/) |
| Does the model predict withheld stars? | All25 spatial folds complete:49,541 fixed held cores across five fields. Bright-star training/prediction gaps recur; the scores are mean normalized squared residuals, not dof-corrected reduced chi-square. [K5 audit](review/k5_aggregation_review.json) |
| Is the fitted solution unique/stable? | Alternate training-only spatial start gives similar overall scores but nonzero flux/WCS differences; both endpoints reach the final cap. No precision-level route equivalence or full convergence is established. [Start control](crossfit/start_control_report/) |
| Is the feature physically intrinsic to the ePSF? | RAW/prepared-image and neighbour-basis controls are attribution-inconclusive; extra nuisance flexibility also absorbs independent mismatch. [RAW controls](raw_control/) |
| Are external faint fluxes physical truth? | Faint calibration supports conditional stress tests, not fixed-flux or absolute-bias acceptance. [Calibration](calibration/extension_C4/) |

## Final confirmation

The reserved exposure was selected using metadata before its model outcomes were seen. Fresh header/epoch geometry, catalogue roles, analytic seed and strict training-only initialization replace development-image geometry. The prospective three-family contract fixes source support, endpoints, two baseline contrasts, common block draws and97.5% intervals before unblinding. All three fits and the locked evaluation are complete. The independent reviewer reproduced both bootstrap columns exactly and inspected the final residual images. [Protocol](review/CONFIRMATION_RESERVATION.md) · [Inputs](confirmation_geometry/README.md) · [Training audit](review/confirmation_training_review.json).

## Provenance and reproduction

Each linked subreport records exact commands, immutable code/input hashes, source IDs, masks, figures and inspection. Scientific products remain under this `/astro` directory. Shared live production code was not edited. Numerical/reference qualifications are reviewed in [SIGNOFF.md](review/SIGNOFF.md). Owned source/doc snapshots are committed and pushed on separate development branches; no production-model integration is implied.

The central experiment index records completion with limitations. The final result is a negative closure conclusion with a supported reference-recipe choice; completion must not be confused with proof of zero residuals.

## Historical execution notes (superseded by the final conclusion)

### Prerequisite recovery, 01:56 EDT

Existing atlas recovery watcher was stuck parsing empty condor_q output after successful trial exit. Independently verified trial 2377.0 exit0 and validated recipient/closure report, then continued exactly its planned recovery: existing DAG345 jobs2401.0 and2419.0 changed16GB→32GB and released. All source, inputs and job identities unchanged. Evidence recovery/guarded_recovery.json. Existing watcher remains untouched and guards against repeated release.

## Mid-investigation checkpoint

Five saved full A3 states have been replayed on original scene_boot with pinned renderer3609d46; all are in-sample, not final validation. All original stage3 fits reached cap. Recent neighbour A/B/C pilot is chroma_model=none and cannot directly certify full A3 colour performance. F1 linear extraction fitted-noise null is numerically reproduced, but preliminary mean-stack rejection is not intrinsic-ePSF attribution; robust influence checks and source-centered injection support review pending. Independent source/noise/model reviewers remain active. Existing C4 ledger campaign continues; both recovered jobs exited0.