# Full-model inventory and numerical closure, 2026-10-04

Status: running. Owner: model_campaign under epsf_model_closure_20261004. This work preserves the separate atlas/neighbour campaign and its watcher.

Question: what residual magnitude/colour structure remains in the actual full spatial/chromatic A3 model, and is an extension identifiable after numerical convergence and scene accounting controls?

## Input and implementation contract

The existing cell-6 neighbour pilot explicitly disables chromatic terms, freezes WCS and anchor/faint templates, and trains one local ePSF. It cannot establish closure of full A3. This campaign first replays the five all-star fits in `/astro/armin/koji/syndiff/dev_runs/paper_dataset_20261001/{C4,F1,F2,S22,C1}/fit` on their saved `scene_boot`, `flux_solved.npz` state and `params.npz`. These are archived F=4 bootstrap hp_d inputs; omitted faint-neighbour light remains an input limitation. The ledger owner audits light accounting; a ledger publication is not assumed to change image pixels.

Runtime: immutable task-local git archive of production `3609d467a7f17c9ed56a434d71f131ca5cf15528` in `pin/`. C4 producer is `81bc761055be623e92a892b5f768df5ccfa9d0e1`; renderer code is identical (the intervening forward_model change is only chain/scene.py). Other four producer fits record 3609d46. Diagnostic sources live on isolated dev branch `codex/epsf-model-closure-20261004`, worktree `/home/kshukawa/syndiff_pipeline/dev/.claude/worktrees/epsf-model-closure-20261004`. Input file SHA256 hashes are written per field; no stored model is silently decoded/re-encoded.

A3 uses a 6x6 grid of 63x63 raw ePSF nodes and 17 colour coefficients, with Gaia-XP u and recorded BP-RP fallback, raw generator gauge and global reference colour. The coefficients are radial shift degree 2; radial kurtosis degree 1; global blur; global dilation; axis trefoil; radial dilation; centred colour-quadratic radial shift degree 1; and two detector elongation planes. The saved fits have no brightness-width term, no fitted background, lp7s strength 3e8, no adaptive rejection. Fitted per-star source amplitudes and polynomial WCS are distinct from profile shape terms.

An extra linear-in-colour round-width coefficient would duplicate A3. It is excluded from the candidate list. Previous `colourwidth_r2_20261001` tested A3, fixed brightness width, radial blur/dilation extensions and combinations on F1/F2 folds0/1; neither radial form passed its rule and stopping times varied. This campaign does not repeat that ladder.

## Execution and validation

Activate syndiff, then run `code/render_baseline.py FIELD` with `OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 taskset -c 0-7`. Exact logs and `/usr/bin/time -v` are under `logs/`. Five fields run serially. The C4 first render took 20 seconds to compile/evaluate and 2.8 GB peak before a diagnostic exporter JIT-context mistake; the adapter was fixed without changing runtime/model and replay restarted. This was a diagnostic failure, not a fit failure.

Native outputs preserve data, full scene model, residual, noise, source templates, source IDs, repeated-pixel UIDs and ownership, masks, active support, saved flux priors/free flags, catalogue Tmag, Gaia BP-RP and fitted XP-u colour. Saved flux/chi-square parity is measured. Standard compatibility rasters keep legacy BP-RP and fitted-flux-derived Tmag definitions explicitly; primary scientific grids use catalogue Tmag and fitted colour u. These quantities are never substituted without a label.

All five archived fits end at stage3 step1999; their last gradient norms range13–33. A DONE file therefore does not demonstrate convergence. The first numerical test optimizes existing colour coefficients on fixed ePSF/WCS, checks actual-renderer directional derivatives and generator rank, before any new family is trained. It is a conditional numerical control and cannot replace full joint convergence, independent starts or held-out selection.

## Decision status

No new term selected. No noise-limited, model-complete or improved-photometry claim. Current full A3 residual images, fixed-support comparisons and numerical quality are being measured. Final confirmation needs supported input generation, separate development/confirmation stars, fixed evaluation support, K=5 training-only initialization and calibrated impact/structure controls owned by the parent investigation.

Conclusion (2026-10-04): archived full spatial/chromatic models are available on all five fields, but the one-cell neighbour pilot is a different model and current five-field science K=5 runs are not present. Numerical convergence remains unestablished. The first required result is exact full-model replay, not an additional redundant colour term.

## Completed development controls and current handoff

The archived replay and 25-image gallery are complete. Independent UID ownership, saved-state and scene reconstruction checks pass. [Matched-population controls](matched/index.html) preserve source role and match magnitude/colour, position, phase, crowding and mask support. They do not support a common magnitude-width correction. Colour-linked asymmetric residuals remain; the own-profile translation/Laplacian/amplitude/background projections retain 22–65% of residual energy. These descriptive in-sample projections do not establish physical PSF variation.

[Numerical controls](preflight/index.html) preserve the intended native estimator. Its nuisance stop-gradient is a deliberate role policy, so its estimating direction need not equal a finite difference of the moving-nuisance scalar objective. Matched native/frozen Adam 100 controls give almost identical colour coefficients. The fixed-nuisance L-BFGS run terminates ABNORMAL and no stationary-fit claim is made. [Template-only response controls](template_response/index.html) on C4/F1 retain the visible population residual patterns, with all support/noise/priors and starting model held fixed; they precede the separate full-background rebuild.

The bounded candidate is now A3 versus A3 plus two CCD-coordinate colour-translation coefficients, not another redundant round-width term. [Construction checks](candidate_preflight/index.html) establish exact zero-initialized parity on a shared adapter graph and numerical nonredundancy. The later strict-fold training-only check includes 17 colour, 2 added shift and 42 polynomial WCS directions: rank 61, normalized condition 33.16 on 50,000 training pixels; added shifts retain 52.8% and 63.4% norm after nuisance projection. The full ePSF base/mode tangent is not projected, and physical attribution still requires independent evidence. No reparameterization was selected using confirmation outcomes.

All five latest full input generations are now ready under `inputs/full_generation/<FIELD>/scene_boot`, with newly built native-template background and latest F=4 Hotpants preparation. Every original scene array except the observed data is preserved, including noise; `inputs/manifest.json` carries hashes and producer records. The parent crossfit owner has the frozen candidate adapter and owns full joint fits, held-out comparisons and the separate confirmation exposure. These replace the earlier missing-K5 status as work proceeds; this subsection itself makes no completed-K5 or accepted-model claim.

Conclusion at this handoff: preparation and numerical controls are complete enough for the bounded A3-versus19 comparison on the new full inputs. Neither noise-level closure nor a preferred extension has been established by this model-development package. Parent-held independent scoring, confirmation and impact tests determine the final recommendation.