# Template-only response controls, 2026-10-04

Owner: model_campaign. Status: four bounded native17 controls completed; figures being inspected. These are explicitly separate from the current full-background generation and the independent fold comparisons.

For C4 and F1, the old archived bootstrap scene and a new template-only hp_d use exactly the same source IDs, pixel UIDs, valid/owner masks, noise, roles, island graph, saved flux priors and fixed ePSF/WCS. Only the observed image differs. The template-only images use the latest F=4 template with the original background. Input owner records are ../inputs/template_only/{C4,F1}/DONE.json and support_check.json. The later full_generation products rebuild the background too and supersede template-only for the main new-model comparison.

Each old/new fit begins from that field's identical archived all-star params.npz, frees only the original 17 colour coefficients and performs 100 native Adam steps at learning rate 1e-4. Native stop-gradient source roles are unchanged. Code: fit_native17.py on codex/epsf-model-closure-20261004, runtime git archive3609d46. Exact CLI logs and resource measurements are ../logs/{C4,F1}_{old,new}_native17.log. Fixed step cap is a numerical response control, not convergence. Independent confirmation is owned by the parent crossfit task.

template_response_figures.py uses common stars and support in both arms, catalogue Tmag and fitted colour u for the primary comparisons, original catalogue-centred registration, and the archived fitted amplitude as a common normalization. It shows observed data, full fitted models, residuals and new-minus-old stacks. The legacy standard grids keep BP-RP and baseline-fitted-amplitude Tmag definitions. Own-image objective/chi-square values refer to different observed images and cannot select a model family.

Conclusion: these controls measure the response to changing the template while holding the background and evaluation contract fixed. They cannot certify the full new input generation, an intrinsic magnitude effect, a preferred model family, independent photometric accuracy or a stationary fit.