# Independent implementation review — 2026-10-04

Status: FINAL independent review complete, 2026-10-05 UTC. Retain A3 for the tested conditional recipes; do not promote CCD19 or the fixed-width extension. Noise-level closure and unique convergence are not established. The final sections and FINAL_CHECKLIST.md supersede historical pending-status statements preserved below.

## Historical F1 fitted-noise map

Reviewed `/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/fitted_null.py` against the installed `syndiff_pipeline/star/band_joint.py` solve and archived raster helper. Its nuisance SVD rank rule and target projection match production. Per-star noise propagation is correct:

$$A=\sum_i B_i D_i(I-Q_iQ_i^{\mathsf T})S_i,\qquad V=A A^{\mathsf T}.$$

Here $i$ indexes stars; $Q_i$ contains orthonormal weighted nuisance and target columns (dimensionless); $D_i$ is the diagonal matrix of supplied image-pixel standard errors in image units; $S_i$ selects that star's pixels from a vector of independent unit-normal variables indexed by physical FFI pixel; $B_i$ combines rasterization, fixed catalogue-flux normalization, group membership and fixed mode weights, with reciprocal image units; $A$ maps shared standardized noise to all dimensionless mode amplitudes; $V$ is their dimensionless covariance. Entries with repeated physical UIDs are added **before** forming $V$. Summing per-stamp covariances independently would omit cross-terms.

Independent reproduction uses three actual overlapping targets 139, 928 and 574, selected only by position: separations from the first target are 0.355 and 0.969 pixels. Original raster helper acting on unit pixels builds the map independently. Twenty shared-noise realizations are separately refitted by the installed production solver, with no import or execution of the significance script.

- 1,173 stamp observations collapse to 410 physical-pixel noise variables.
- Maximum direct-refit versus summed-operator discrepancy: $3.53\times10^{-17}$ in the dimensionless dipole functional.
- Correct shared-pixel variance is 1.25887 times the naive independent-stamp variance for this particular functional; the ratio is not universal.
- An independent 100,000-draw simulation gives empirical/analytic variance 0.99847.

Code and full numbers: [check_shared_noise.py](check_shared_noise.py), [shared_noise_review.json](shared_noise_review.json). The data underlying this check are shown below, with original masks and equal data/model scales in each row. Figure visually inspected 2026-10-04. This is numerical estimator validation, not a finding that the original residual is Gaussian.

![Three overlapping data/model/residual stamps](shared_noise_check_stamps.png)

**PASS:** conditional linear refitting, shared-pixel covariance propagation and Gaussian feature simulation. **NOT ESTABLISHED:** supplied error-map correctness, independent detector noise, upstream learned-model uncertainty, population-representative PSF truth, or absolute flux/centroid accuracy.

The declared family comprises six modes across retained marginal magnitude, colour and detector-cell groups. Global maximum standardization of all 84 correlated features provides the requested within-family correction under the conditional Gaussian reference. The full colour × magnitude × cell interactions are not included and cannot inherit a closure claim. Excluded detector cell 8 and unsupported bins remain unmeasured. The already viewed historical sample is development evidence; predeclaring the new code's statistic does not retrospectively make the underlying data blind. It also does not correct earlier exploratory choices of this diagnostic.

Review requests sent to significance owner before interpretive signoff:

1. Report null exceedance counts, simulation resolution and Monte Carlo bounds. Zero exceedances among 100,000 draws does not measure an exact tail probability of $10^{-5}$; the plus-one value is a simulation convention.
2. Show underlying cell-1 data/model/residual images next to the extreme global statistic; improve saturated magnitude-grid data/model displays without hiding residual scales.
3. Construct flux-conserving perturbations on the full finite source profile before extraction-stamp clipping. Actual integer shifts of up to one pixel occur in the selected cohort. Alternatively label the current power controls as finite-extraction-stamp zero-sum defects and quantify their boundary sensitivity. First-order width/displacement perturbations are not an exact broadening/translation at large amplitude.
4. Preserve the distinction between rejection of supplied pixel-only noise and an intrinsic common ePSF error: conditional mean-stack rejection is highly sensitive to contamination/outliers. A robust alternative requires its own refitted null rather than borrowing the mean covariance.

## Full A3 replay / native model export

Reviewed `models/code/render_baseline.py`, its pinned `crossfit/score_oof.py`, and pinned `scene_fit.py` templates/diagnose. The exporter reconstructs the saved source state before calling the actual renderer and does not renormalize individual templates. That is the correct replay contract. Saved-source nuisance stop-gradient affects learning, not this forward value.

Independent code [check_model_export.py](check_model_export.py) sums the exported template responses times re-solved amplitudes by physical pixel UID in float64, without importing the renderer. Across C4, F1, F2, S22 and C1:

- Source ordering and all four saved state arrays match exactly.
- Valid repeated UID observations have exactly identical data; every unique valid pixel has total ownership one.
- Independent full-scene reconstruction RMS is $5.34$–$5.75\times10^{-6}$ times supplied pixel noise; maximum discrepancy is below $4.76\times10^{-4}$ times supplied noise. This is compatible with float32 forward/export rounding at the stated diagnostic accuracy.
- C4's largest replay flux difference relative to $\max(|F|,1\,e^-\,s^{-1})$ is $4.97\times10^{-6}$; its 99th percentile is $1.12\times10^{-7}$. Other fields' maxima are no larger than $2.37\times10^{-7}$; S22/C1 amplitudes match bitwise.

Here $F$ is the archived fitted amplitude in image $e^-\,s^{-1}$ units. The one-unit floor is a numerical comparison convention for near-zero amplitudes, not a photometric acceptance threshold. Full results: [model_export_review.json](model_export_review.json). These numerical checks do not certify convergence or held-out quality. The model owner supplies the native residual grids underlying scientific conclusions; no new shape claim is made here.

**PASS:** exported arrays reproduce the pinned saved state and represented model at negligible numerical error for residual diagnostics. **CAVEAT:** template sums are not exactly one (C4 median 0.99827, full range 0.97375–1.07030); integrated native-stamp flux differs from the amplitude parameter. Preserve the actual renderer convention and never silently renormalize it in a comparison, calibration or injection. Determine full-profile normalization and phase/gauge conventions before an absolute bias interpretation. Standard legacy raster bins use fitted-flux magnitude and BP−RP; primary scientific tests must use separately labelled catalogue magnitude and actual fitted XP-u colour.

**Conclusion:** independent checks validate the conditional shared-noise mathematics and archived full-model replay. They do not validate a completed scientific model. Full significance interpretation awaits the requested caveats/figure corrections and independent covariance/model controls; calibration, numerical convergence, family selection and untouched confirmation remain open.

## Native nuisance stop-gradient policy — independent interpretation

The pinned `scene_fit.make_model.templates` deliberately applies `jax.lax.stop_gradient` to role-2 nuisance templates. Their forward values nevertheless come from the current model parameters at each call. This is explicitly documented native role exclusion; ordinary finite differences of the displayed scalar loss are therefore not the same derivative as the implemented update field.

$$g(\theta)=\left.\frac{\partial\mathcal L(\theta,\bar T_n)}{\partial\theta}\right|_{\bar T_n=T_n(\theta)}.$$

Here $\theta$ denotes all learned model parameters in their native units, $T_n(\theta)$ is the current nuisance-template response array, $\bar T_n$ is a separate fixed array for a single derivative evaluation, $\mathcal L$ is the reported scalar objective with a differentiable flux solve, and $g$ is the implemented update direction with units of loss per parameter. In contrast, ordinary finite differences of the fully refreshed forward function also include the term through $dT_n/d\theta$. An AD/finite-difference mismatch is expected for this deliberately modified derivative rule. It does not invalidate the archived forward replay, and does not on its own identify a coding error.

Three policies are scientifically different:

| Policy | Forward nuisance profile | Direct nuisance derivative | Interpretation |
|---|---|---|---|
| Native dynamic stop-gradient | Recomputed at every model value | Suppressed | Role-excluding estimating equation; not the total derivative of the displayed refreshed objective |
| Frozen nuisance at common start | Fixed throughout optimization | Zero because fixed | Coherent scalar objective conditional on the frozen nuisance profile |
| Full differentiation | Recomputed | Included | Nuisance sources directly train the model; changes role contract |

Do not silently patch a derivative test until it passes by changing the model policy. The minimal fair diagnostic compares native and explicitly frozen-nuisance arms at identical initial forward arrays, support, source flux policies and nominal schedules. Record policy in output metadata and compare actual independent residuals/flux response and update stability. A native update zero is an estimating-equation solution, not necessarily a local minimum of the displayed refreshed objective; a frozen-arm scalar decrease is conditional on that frozen shape. Fixed-nuisance AD/finite-difference agreement can isolate numerical differentiation quality, while the native rule requires fixed-array directional differentiation or an estimating-equation stability analysis. Full differentiation is a distinct scientific arm, not a routine repair.

Conclusion for this gate: **retain original policy as recorded; permit bounded, separately labelled policy controls; withhold large-campaign acceptance until the intended objective/update contract and held-out response are resolved.**

## Updated F1 interpretation and robust endpoint review

Reviewed `significance/robust_null.py` and the revised full-source injection construction. The latter now builds zero-sum perturbations on each source's full stored profile before transporting to the physical FFI grid; the extraction-stamp clipping concern is resolved. The controls remain first-order perturbations of the fitted profile, not independent empirical PSF truth. Current centroid power is x-direction only; y-direction power was requested because the strongest median feature is a y dipole.

For the median endpoint, each shared Gaussian pixel draw is propagated through the exact per-star residual/raster map **before** taking the median. An independent pilot estimates feature standard errors; separate draws set the maximum-statistic reference and check its rejection fraction. This correctly accommodates the median's nonlinearity under the stated fixed-model null. A column-common Gaussian term is generated using physical detector column IDs; it is a conditional covariance sensitivity, not an upper confidence envelope for all detector/systematic covariance.

The historical F1 median result is inconsistent with the supplied pixel-only fitted-noise reference: maximum standardized feature 10.630 versus global 95% reference 3.455, with 0/2,000 simulated maxima as large (one-sided 95% Monte Carlo upper tail bound 0.001497). The independently motivated column-common sensitivity remains inconsistent, 10.639 versus 3.481, also 0/2,000. These numbers concern a deliberately limited conditional simulator; they are not physical ePSF-only discovery probabilities. The median endpoint was added after seeing the mean rejection, so it is exploratory robustness evidence on already viewed data, not untouched confirmation or an originally predeclared combined mean/median family.

![F1 robust median stacks and noise-reference bands](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/robust_null_images.png)

Inspected 2026-10-04. Bright residuals have an opposed central positive/negative feature; the shape is not a pure radial width. The underlying data, fitted model and residual accompany the mode plot. Bands show the null simulator's fluctuation scale; they must not be interpreted as generic simultaneous confidence intervals on heterogeneous physical shape errors.

The much larger mean statistic is dominated by a small number of strongly structured source stamps. The strongest cell-1 core functional has 65.1% of its absolute individual contributions from one target and 95.8% from the top five; this is influence concentration, not the fraction of signed signal removed by deleting them.

![Actual cell-1 mean and influential source images](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/mean_influence_images.png)

Inspected 2026-10-04. Individual targets 488 and 681 show substantial vertical column structure. The figure now shows the images behind the global mean statistic. No residual-value clipping or source rejection was used to obtain the median endpoint. The result rules out interpreting the mean rejection alone as a typical intrinsic ePSF width morphology. It does not identify which detector/preparation/PSF mechanism produced either pattern.

**Updated conclusion:** F1 residuals are not explained by the tested fixed-model pixel-noise simulator, even with the tested column covariance. A weaker central structure survives a median statistic, so dismissing all evidence as a few outliers is also unsupported. Absolute calibration, broad unknown covariance, independent profile truth and cross-field confirmation remain open. These conditional results justify further attribution and a supported model comparison; they do not certify any final model or universal magnitude response.

## Strict-fold mask review

Independently checked the saved F1 fold-0 strict scene against its unmodified archived base. [strict_mask_review.json](strict_mask_review.json) records the scene hash and exact checks. Source IDs, UID mapping and pixel-centre geometry are unchanged. Every held image pixel occurrence has zero data/valid/finite and unit placeholder noise; every represented non-held pixel retains exactly the original data, noise, valid and finite entries.

Reading the pinned solver confirms that nuisance prior precision depends on `finite/variance`, so zeroing only the likelihood `valid` mask would have left a derivative/support path. The wrapper correctly masks both. The strict initializer supplies a sanitized image and adds the same held-pixel mask to extraction; its output parameter dictionary explicitly contains the newly trained photutils base and WCS, with zero chromatic/mode coefficients. It does not copy learned parameter leaves from the all-star A3 checkpoint.

**PASS:** direct prepared-image pixel exclusion in the saved strict scene. **CONDITIONAL:** source lists, roles, retained-light preparation, catalogue transforms and original scene geometry/pooling centres remain fixed upstream inputs. This is not a fully independent raw-to-science pipeline split. A held-pixel perturbation invariance check through preparation/initialization was requested. Parent paused large strict-fold training while the native versus frozen-nuisance estimating-policy contract is resolved; the mask pass does not authorize treating a short preflight as numerical convergence.

## C4 bright-calibration audit

Independent [check_calibration_sources.py](check_calibration_sources.py) maps the **original** `inputs/photometric_calibrators.csv` directly to the saved C4 fit/source arrays by exact integer source ID, independently of the calibration agent's joined audit. All 37 IDs map, the CSV fitted flux agrees to maximum relative $5.08\times10^{-8}$, and weighted four-band flux reconstruction agrees to $1.79\times10^{-16}$. There are 36 role-1 anchors and one role-2 nuisance, all marked free. The nuisance is `1420150560129293440`; “free” does not mean its fitted flux is data-determined when its response is almost entirely masked and its Gaussian prior dominates.

The giant magnitude trend is not explained by substituting dimensionless `scene.tess_flux` for measured flux. Ordinary natural-log slope decomposition for the retained 35 original calibrators gives $+0.00410$ per magnitude for fitted/Gaia-proxy flux and $+0.39618$ for PS1/Gaia-proxy flux, hence $-0.39208$ for fitted/PS1 flux. This decomposition uses a common catalogue proxy for algebraic attribution, not independent flux truth. Excluding the prior-dominated source still leaves a large discrepancy (36-source univariate slope $-0.52015$ per magnitude); this is a different selection from the retained-35 robust cut, so those slopes must not be conflated. The calibration agent's joint robust colour+magnitude regression is also a different estimator, not an arithmetic disagreement. [Exact audit](calibration_source_review.json).

![C4 calibration amplitude, units and catalogue comparison](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/calibration/figs/c4_calibration_units.png)

All four calibration stamp pages were inspected. Many have visible structured residuals or masked regions. Page 2 directly exposes the nearly completely masked nuisance source at its top left:

![C4 calibration stamps including prior-dominated source](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/calibration/figs/c4_calibrator_stamps_2.png)

Other underlying inspected pages: [1](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/calibration/figs/c4_calibrator_stamps_1.png), [3](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/calibration/figs/c4_calibrator_stamps_3.png), [4](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/calibration/figs/c4_calibrator_stamps_4.png). The catalogue band discrepancies and archived conservative bright-limit violations support a catalogue-quality concern, but they do not establish saturation or a unique association failure without independent flags/image evidence.

**Conclusion:** the old C4 bright scale is not a validated faint-neighbour flux calibration. A source-ID/unit/export error does not explain its large magnitude trend, and prior-dominated/masked calibrators must be excluded before any revised empirical calibration. Treating catalogue faint extrapolation as truth remains a failed acceptance gate; there is no justification for absorbing this uncertainty by broadening the residual-noise null.

### Scope correction for the derivative-policy table

The coherent frozen-nuisance objective and derivative checks above apply to the **17-coefficient chromatic-only control with ePSF/WCS fixed**. `loss._epsf_params_for_context` also suppresses ePSF base/mode derivatives for role-1 anchors while retaining their chromatic derivatives. In a full joint fit those anchor forward values can change through the blocked leaves. Freezing nuisance templates alone therefore does **not** remove every moving stop-gradient surface from the full recipe. The displayed equation isolates the nuisance mechanism with other barriers fixed; a complete joint estimating-equation contract must include both anchor and nuisance role exclusions. Do not extend the 17-dimensional finite-difference pass to the full model or silently remove anchor barriers to force agreement.

## Bounded policy-control outcome and continued baseline decision

Reviewed the completed `models/preflight/paired_cross_evaluation.json` and inspected both paired history and residual figures. The 100-step Adam endpoints differ by only $2.60\times10^{-7}$ in chromatic-coefficient Euclidean norm. Evaluating either endpoint with the same nuisance forward policy gives almost identical losses; changing the nuisance forward policy at either endpoint gives about 0.0242 supplied pixel-noise units RMS in predicted image. The near-equality of learned parameters despite distinct scalar loss histories is consistent with the intended estimating-rule interpretation.

![Paired native and frozen nuisance histories](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/models/preflight/paired_histories.png)

![Underlying paired chromatic residual stacks](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/models/preflight/paired_residual_stacks.png)

These are in-sample, fixed ePSF/WCS, chromatic-only controls with a 100-step cap. They support retaining the **original native estimating policy** for the conditional strict-fold baseline; they do not require a silent role-policy change. They do not prove full joint convergence, absence of path dependence or scientific model completeness. A caption clarification of whether the displayed endpoint stacks use the same forward policy was requested; the cross-evaluation table explicitly separates that issue.

The final independent decision for this numerical gate is **PASS to continue the native baseline under its documented estimating-equation contract**, not PASS for final model acceptance. Full-training convergence should assess stability of predictions and estimating equations with independent starts/schedules and held-out data, rather than demand that this deliberately modified update equal the total derivative of an unrelated refreshed scalar function.

## F1 centring and directional positive controls

Inspected the final deterministic centring control and y-direction injection panels. Replacing the frozen WCS stacking origin by WCS plus the renderer's explicit chromatic translation leaves the bright opposed central residual visible; its y-dipole mode changes from $2.794\times10^{-4}$ to $2.958\times10^{-4}$ of fixed predicted total flux per sampled pixel. This is descriptive: no original-grid null covariance is transplanted to the new raster. The native data, fitted model and residual remain unchanged. The shift control therefore does not remove the observed structure or establish an intrinsic PSF mechanism.

![F1 fixed-origin centring control](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/centering_control.png)

The missing y-direction positive control is now supplied: for this finite-support, first-order source displacement and the stated column-common noise model, detection fractions over 500 draws are 6.4%, 39.2% and 100% at 0.001, 0.003 and 0.01 TESS pixels. A 100% simulated fraction is not certainty about arbitrary physical defects. The displayed underlying injected profiles and post-fit median images visibly distinguish radial from x/y dipole modes.

![Full-source mismatch profiles and post-fit median images](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/positive_control_images.png)

**Conclusion:** the conditional numerical and diagnostic implementation gates are now substantially supported. Historical F1 retains a central residual beyond the tested fitted-noise reference, and C4 retains an unresolved catalogue calibration problem. Continue the originally authorized baseline/corrected-input investigation without claiming scientific closure or changing role semantics merely to force a derivative test to pass.

## First strict F1 fold: independent support/scorer review

The first native-estimator F1 fold is now a completed **single-fold conditional shape check**, not just a mask preflight. Independent reconstruction from original source role, physical pixel coordinates, exact fractional-centre radius-three core, original valid mask and fold assignment reproduces all 2,062 exported source IDs. Every scored valid core pixel belongs to the held physical region; none enters the training pixels. Eleven sources near the CCD edge require treating nonexistent exterior pixels as absent, not as an unheld observed region. [Independent check](strict_final_core_review.json).

The held-input poisoning test makes sanitized image and noise planes exactly invariant with fixed source metadata and geometry. It is an input-plane invariance result, not a repeated independent optimizer run. The scorer keeps learned shape/WCS fixed but re-solves evaluation amplitudes **and** recalibrates the global nuisance-prior scale using evaluation contributors, including held stars. It therefore measures conditional shape quality after nuisance fitting, not independent flux prediction or zero-point accuracy.

The initial stage-versus-final residual images used each state's own fitted-flux denominator. Reviewer requested a shared denominator; the owner regenerated the primary comparison using each star's fixed final flux for both states, with the same 2,062 stars. The held-core chi-square curves were already independent of this display normalization. The cap endpoint was not chosen using the held-out scores. An optimization learning curve does not establish final model correctness; the final stage still reaches its iteration cap, and no completed K=5/independent-start claim follows.

![One strict F1 fold: fixed-support learning curve and common-denominator residual stacks](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/crossfit/F1/fold0/diagnostics/learning_and_stacks.png)

The replacement primary image was inspected. Residual structure becomes visibly smaller during fitting on training pixels, while the final held stacks retain central structure. This supports a functioning strict-pixel holdout implementation and improved conditional prediction through the training sequence. It does not select a new model term or certify noise closure.

**Conclusion for this milestone:** the first actual strict-fold support and scoring contract passes independent review with the stated nuisance-refit and prepared-input limitations. Continue the remaining predeclared baseline confirmation and corrected-input work; final model, convergence and scientific-impact acceptance remain open.

Caption clarification from the producing-code audit: the paired Adam residual rows use each arm's own forward policy (native refreshed nuisance versus cached nuisance), so their difference includes both parameter and forward-policy effects. It is not solely a parameter comparison. The 2×2 cross-evaluation table separates those effects; the nearly white averaged difference at the plotted scale does not contradict the nonzero physical-pixel RMS quoted above.

Current reviewer milestone conclusion: numerical diagnostic, replay and single strict-fold contracts pass within their explicit conditional scopes. Final scientific acceptance is still open pending the remaining calibrated-input, convergence, model-comparison and confirmation evidence.

## Static seed and reference-profile audit

Read the actual pinned `Scene.contexts`, `make_model`, `StaticContext`, `forward_model` and initialization order, including the optional local-cache renderer. Native `scene_fit` passes no local profile cache: every role's template comes from the current parameter dictionary. Anchor and nuisance stop-gradient rules suppress derivatives, not refresh of forward profile values. `scene.z['epsf_seed']` and source-bundle `params0` provide the default initializer, which an explicit `--init-params-file` override replaces before rendering. Contexts contain geometry, catalogue features and grid locations, not an old shape bank.

The independent in-memory poisoning test [check_static_seed.py](check_static_seed.py) sets the old scene seed, all floating source-bundle `params0` leaves, and both stored `epsf_base`/`epsf_modes` banks to NaN. After loading the actual training-only photutils initializer, reconstructed contexts render three sources per role with **bit-exact unchanged templates** for contributors, anchors and nuisance. Changing the current ePSF base leaf changes templates in every role, which is the positive control. No saved science input was modified. [Full result](static_seed_review.json).

![Actual role templates under static-seed poisoning and current-leaf positive control](static_seed_images.png)

Figure inspected. **PASS for the native scene-fit path:** replacing the explicit initializer does not leave an old hidden forward shape bank. This does not certify a different packed/cache-enabled renderer, and does not erase any old image-derived source geometry/selection. The prospectively reserved exposure still uses fresh header/catalogue construction and training-only initialization; its static-bank policy is documented in [CONFIRMATION_RESERVATION.md](CONFIRMATION_RESERVATION.md).

## Candidate non-redundancy with polynomial WCS

Independently recomputed the column-normalized Gram eigensystem and Schur-complement projections from the saved arrays. The old conditional 19-colour-column check gives rank 19 and condition 40.0023. The newer strictly training-only stage-1 check includes the 17 existing colour columns, two proposed CCD colour shifts and all 42 polynomial-WCS coefficients: rank 61, normalized condition 33.1627. The added x/y response norms retain 52.78% and 63.42% after projecting out the other colour/WCS columns. All 50,000 sampled points are unique, valid, owned, active physical training pixels and lie outside the held-pixel mask; this was checked independently from the stored scene and sample indices. [Audit](candidate_rank_review.json).

![Training derivative Gram evidence](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/models/candidate_preflight/joint_wcs/identifiability.png)

![Actual new-full training residual and added-shift derivative images before/after colour+WCS projection](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/models/candidate_preflight/joint_wcs/derivative_stamps.png)

Both figures inspected. This establishes a non-redundant **conditional** direction at this training checkpoint with base/modes fixed. It does not project against the full ePSF-shape tangent, establish physical chromaticity or predict held-out benefit. The family is unchanged; no new data-guided reparameterization was introduced.

## First matched A3–CCD19 development result

Read `crossfit/code/eval_joint.py`, inspected the F1 full-generation matched residual image and independently recomputed the primary statistic and **all 10,000 bootstrap draws bit-for-bit** from `paired_scores.npz`. The result uses 2,062 fixed held cores in 49 dependence components, with the largest component containing 4.90% of the stars. The median paired log score ratio is −0.00011749, with conditional 95% interval [−0.00104458,+0.00053577]. [Independent arithmetic](F1_development_pair_review.json).

![Full-generation F1 development pair; same physical cores and flux denominator](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/crossfit/F1_full_generation/fold0/paired/paired_residuals.png)

The changes are visibly small while common residual structure remains. **No primary improvement of CCD19 is established on this F1 development fold.** The interval is conditional on fitted models and realized evaluation-calibration scales and does not include all global calibration/noise uncertainty. It is not an equivalence claim or proof that colour shifts never matter in other fields. Both fits are cap endpoints, and this exposure had already informed development.

Reviewer found a minor safety-table implementation mismatch: fractional one-third CCD cell boundaries versus the frozen integer 683/1366 boundaries. It affected two x and one y assignments, not the primary statistic or bootstrap. The owner corrected the code and safety table. The independent actual-leaf ablation is a useful implementation check; an improvement over zeroing a leaf at the candidate's other fixed parameters is not a substitute for the fully refitted A3-versus-CCD19 comparison.

**Current conclusion:** native static-profile reuse is cleared as a hidden-shape-leak concern for this path, and the proposed extra colour shifts are conditionally non-redundant. Their first actual F1 matched development comparison nevertheless has no established primary gain. Retain A3 as the simpler supported baseline for that comparison; await the already defined remaining field and prospective confirmation evidence before making a final family decision or noise-closure claim.

## Actual strict-fold conditional residual reference

Reviewed the original-generation T8–13 strict F1 scorer in `significance/strict_fold0/score.py`. Its sparse native-template solve preserves the renderer normalization, ridge, finite-mask nuisance precision and shared physical UID data. The covariance expansion is algebraically correct because the measurement weights touch only valid scored pixels; direct/shared-image and adjoint checks pass. The largest native-versus-float64 mode discrepancy is only 0.0104 conditional standard errors. The separate nonlinear median-prior two-pass simulation reproduces the actual evaluation scale-estimation rule, with explicitly limited simulation precision.

The 300-feature reference comprises 50 supported covariate groups and six native-pixel modes. It uses its own joint covariance/reference, not the historical 84-feature threshold. The observed maximum 244.044 versus 3.7603 rejects the stated fixed-profile supplied-noise reference; this must never be phrased as a physical “244-sigma” ePSF detection. The learned shape/WCS and their relation to training noise remain fixed, so this is not a full train-and-refit null. Actual generator controls show sensitivity conditional on finite native support and fitted-profile derivatives, not independent physical PSF truth.

![All supported original-generation strict-core residual images](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/strict_fold0/all_group_images.png)

![Influence and descriptive median control](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/strict_fold0/influence_images.png)

Those figures, the reference-statistic figure and the injected/post-fit power figure were inspected. The central excess is not accounted for by a handful of dominant stars; the median is shown only descriptively and receives no borrowed mean covariance. These results concern the **original preparation and original strict A3 fit**, not the current-generation A3–CCD19 comparison.

**Conclusion:** the native strict-core significance machinery passes independent review for its declared conditional estimator check. It supports investigating remaining structured residuals and does not establish intrinsic brightness dependence, physical parameter-bias bounds or final model acceptance.

## Fixed-width construction gate

The prospective family amendment and its multiplicity/selection rule are in [CONFIRMATION_RESERVATION.md](CONFIRMATION_RESERVATION.md). Reviewed the actual fixed-width preparation and scorer adapters. The development charge array is exactly the capped product of two seconds, the immutable strictly trained catalogue-prior amplitude, and the training-only initialized native template peak. It does not use measured held-source amplitudes or normalize a clipped template. Every held source receives a positive uncapped charge. [Independent arithmetic](width_q_review.json).

The stock width q mapper had a real ambiguity: two duplicate source-ID pairs in the legacy F1 scene are different nuisance rows with different native geometry/peak responses. Independent row inspection confirms distinct bundle indices, so scatter by bundle row can preserve the values. The corrected adapter checks ordered source ID, bundle index, row ordinal and geometry, excluding role/mask fields that legitimately change between training and evaluation. It refuses ambiguous row reordering. The fresh confirmation catalogue is separately reported unique by its producing owner, so these legacy rows are not inherited.

The final saved-q versus direct-array and zero-width native flux/chi-square/residual checks are exactly equal. The scorer explicitly restores the recorded width charge array, training reference and generator units instead of relying on the stock scorer that omitted those arguments. The initial stale pending status/hashes were superseded by the final producer packet before handoff.

![Development charge construction and training reference](../models/fixed_width_preflight/charge_audit.png)

Figure inspected. **GO for the historically fixed-b development construction**, with no physical brighter-fatter interpretation or outcome selection. The fresh confirmation must generate its own train-only q file and pass the same identity/units/scoring gate; the development array is not transferable as independent truth. A3w also requires its own full training prefix because the fixed nonzero width changes the forward model from the first iteration.

**Conclusion:** the proposed fixed-width sensitivity can now be evaluated under the explicitly amended three-family contract. Its construction has been checked; predictive benefit, scientific impact and final confirmation remain to be measured.

## Current-generation significance and prospective input audit

Independently compared `significance/current_F1` A3/CCD19 arrays: selected rows, source IDs, core masks, catalogue normalization, native data, labels and the complete measurement matrix are exactly equal. The separate sparse covariance/reference constructions retain their own native template operators. Their tiny negative covariance eigenvalues are numerical roundoff, not a material non-positive covariance. Native replay errors are negligible relative to the reported conditional mode scales. [Checks](current_significance_review.json).

The common residual and zoomed difference images were inspected. Both fitted families retain the central structure; their difference is much smaller than the common residual. The maximum conditional statistics, approximately 244.13 and 244.25, must not be compared as a model-quality ranking or quoted as physical Gaussian sigmas. The declared three-model Bonferroni allocation gives per-arm thresholds approximately 4.04. Its zero-exceedance Monte Carlo upper bounds reproduce exactly; the distinction between simultaneous Monte Carlo-bound coverage and inferential family correction is explicit. The separate paired primary comparison remains the correct relative-performance endpoint.

![Current-generation matched residuals and descriptive differences](/astro/armin/koji/syndiff/dev_runs/epsf_model_closure_20261004/significance/current_F1/comparison_images.png)

A metadata-only independent audit of the **fresh reserved** exposure also passed. All 14,933 scene IDs are unique and match the fresh catalogue through exact bundle-row mapping; all role changes are the frozen policy or common demotions. Full-header coordinates with the 44-column crop offset reproduce stored positions within float32 representation precision. The source geometry FitBundle data and weights are exactly zero. No reserved science-image values, fitted model, residual or held score were read by this audit. [Code](check_fresh_geometry.py), [result](fresh_geometry_review.json).

**Conclusion:** current-generation A3/CCD19 conditional diagnostics and fresh input identity/geometry are independently supported in their stated scopes. Width outcomes, numerical robustness and the reserved three-arm confirmation still determine the final bounded model decision.

## Final five-field and numerical robustness review — 2026-10-05 UTC

The preceding sections preserve the chronological development record. Their pending-outcome statements are superseded by this final section and [FINAL_CHECKLIST.md](FINAL_CHECKLIST.md).

Independently audited all five K=5 aggregations: F1 10,536, F2 8,762, S22 8,491, C1 8,224 and C4 13,528 fixed held cores, totaling **49,541**. Every scored row has exactly one held fold, all scored valid core pixels belong to that fold's withheld support, and the reported in-training comparator is exactly the median of that same star's scores from the other four current fold fits. It is not an old all-star model. Checkpoint identity and scored-source identities pass. The one nonpositive C4 amplitude remains in primary metrics; only its flux-normalized display is exceptional. [Independent audit](k5_aggregation_review.json), [code](check_k5_aggregation.py).

![Actual five-field held data, model and residual images](k5_data_model_review.png)

![Fixed catalogue-normalized residual images by magnitude](k5_residual_review.png)

![Five-field score levels and descriptive held/training contrasts](k5_score_review.png)

All three figures inspected. The bright-bin held/in-training ratio is about 6–16.6 across fields, but spatial coverage, initialization and contributor/anchor role are confounded with magnitude. These are descriptive generalization contrasts, not causal proof of overfitting or intrinsic magnitude dependence. The score divides the normalized squared residual sum by the valid pixel count; it is **not reduced chi-square with an established independent degrees-of-freedom correction**.

The different-start F1 control uses the training-only empirical node average as a spatially constant shape and header WCS, with matched priors/support. Its fine-profile information remains related to the reference initializer. It measures numerical spatial/WCS-start sensitivity, not independent fine-profile truth. For 2,062 fixed cores the median score ratio is 1.00353, with conditional 95% interval [0.99836,1.00958]. Median absolute flux route difference is 0.0592%, and the 95th percentile is 0.2946%. Median/p95 WCS route displacements are 0.00436/0.01819 pixels. These differences are neither astrophysical biases nor equivalence bounds. Both final stages hit 2,000 iterations; convergence is not certified.

![Same-source data/model/residual and differences under two spatial starts](../crossfit/start_control_report/figs/magnitude_images.png)

![Spatial-start flux/score and endpoint sensitivity](../crossfit/start_control_report/figs/endpoint_metrics.png)

These figures and the actual profile/WCS difference panel were inspected. Source evidence is [the start-control report](../crossfit/start_control_report/README.md). A wider independent fine-profile initialization test is not claimed and is not required to conclude that convergence remains unestablished.

## Final current-generation width and noise-reference review

The current F1 fixed-width arm completed its own full prefix with the historical coefficient fixed and additional CCD shifts zero/frozen. Its common-core development log-score contrast is −0.00049915 with conditional 97.5% interval [−0.00309222,+0.00176717]. The corresponding two-colour-shift interval is [−0.00116205,+0.00071523]. Neither establishes improvement. These are separate paired predictive intervals, not comparisons of residual maximum statistics.

![Actual current A3/width residuals and common-denominator differences](../significance/current_F1/width_comparison_images.png)

Figure inspected. Width changes part of the compact central feature but common residual structure remains. Its native/sparse feature discrepancy is 0.01030 conditional standard errors, and training replay maximum absolute log-score discrepancy is 2.61e−7. All three arms have identical 300-feature measurement/support/denominator contracts. Their own maxima (approximately 244.13, 244.25 and 236.01) exceed their three-model-corrected conditional references (approximately 4.04). The reported adjusted zero-exceedance Monte Carlo upper bound is 0.0006141 under the specified simulations; Monte Carlo-bound confidence and inferential multiplicity are distinct. **Do not compare these maxima as predictive quality or call them physical Gaussian sigmas.** Learned-profile, preparation and complete detector-noise uncertainty are absent from the conditional reference.

## Prospective reserved exposure: audit before release and independent outcome reproduction

The frozen metadata-selected F1 exposure, family amendment and decision rule are in [CONFIRMATION_RESERVATION.md](CONFIRMATION_RESERVATION.md). Fresh catalogue/header geometry and analytic seed construction were audited separately from pixels. The strict initializer contains no held sources in its WCS/shape pool; all 835,584 withheld physical pixels are excluded from training. The original-valid geometric cohort of **2,249** ordered cores was saved before outcomes. Harmless compressed-FITS quantization of masked noise placeholders was documented rather than confused with actual held training weight. The fresh width q is reconstructed from this exposure's catalogue, strict training calibration and training-only initializer peak, with exact row/geometry identity; no development or held fitted charge was transferred.

[Training audit](confirmation_training_review.json) and [pre-unblinding audit](confirmation_preunblind_review.json) establish the common initializer/support, final-checkpoint identity, frozen width/CCD leaves and fresh q. The CCD19 resource recovery cold restarted the entire unchanged recipe; no mixed attempt or interim held checkpoint was selected. A single evaluator was released only after those guards passed.

Both primary bootstrap columns independently reproduce **bit-for-bit** on the exact pre-frozen cohort. There are 50 dependence components; the largest contains 3.20% of the sources. The same 10,000 component draws serve both baseline contrasts. [Independent numerical reproduction](confirmation_outcome_review.json).

| Candidate relative to A3 | Median paired score change | Conditional 97.5% interval |
|---|---:|---:|
| CCD19 | +0.375% | −0.225% to +0.905% |
| A3 plus fixed width | +1.865% | +0.712% to +3.071% |

These percentages exponentiate the median paired log score ratio. Positive means worse. The two 97.5% intervals implement the frozen Bonferroni allocation for nominal conditional 95% family coverage; their block-bootstrap coverage is approximate, not a theorem about all detector/model uncertainty.

![Prospective matched residuals: all three arms and magnified differences](../crossfit/F1_confirmation/fold0/confirmation/confirmation_residuals.png)

![Prospective absolute core-score levels](../crossfit/F1_confirmation/fold0/confirmation/confirmation_score_levels.png)

Both figures inspected. Absolute rows share a scale and all use the same A3 fitted-flux denominator; the differences share a separately magnified scale. Central and asymmetric residual structure persists in all three arms. These mean residual images are descriptive, while the declared endpoint uses paired per-source median log scores. Absolute scores are large and strongly magnitude dependent. The common fresh population and initialization differ from development; role transitions also remain confounded with magnitude. No result justifies discarding this prospectively reserved exposure after viewing it.

**The result is a conditional recipe comparison, not converged-family equivalence.** Shapes/WCS remain fixed during scoring, but evaluation fluxes and the global nuisance-prior scale are refitted. Preparation used the whole image. The optimizer cap/scalar-plateau endpoints and substantial native quasi-gradients do not establish unique convergence. The fresh analytic/header seed followed by training-only photutils initialization does not make the whole preparation an independent raw-pixel test.

## Final independent signoff

**Review complete. Retain A3 for the tested recipe and domain; do not promote CCD19 or the prescribed historical width term.** Neither extension demonstrates a primary gain, and the fixed-width recipe is worse in the reserved-exposure comparison. The investigation supports this bounded negative decision, with its reproducible numerical and image evidence. It does not establish noise-level closure, unique convergence, absence of physical magnitude/colour effects, or negligible absolute flux/centroid bias. Residual, calibration, preparation and numerical ambiguities remain explicit limits, rather than grounds for another undeclared family search. All five evidence gates are completed/measured; their noise/convergence acceptance conditions fail or remain unestablished as recorded in [FINAL_CHECKLIST.md](FINAL_CHECKLIST.md). No further experiments are requested by this review.