# Prospective confirmation recommendation — 2026-10-04

This is a metadata-only reservation recommendation to the coordinator, made before inspecting any pixels or model results from the selected exposure. It is separate from the five historical development images and their subsequently strict pixel folds. The latter can measure conditional performance but cannot retrospectively undo the use of those images to choose the model family.

## Reserved exposure

**F1 / S24 C2 K2: `tess2020110032919-s0024-2-2-0180-s_ffic.fits.fz`.**

Absolute input: `/astro/armin/koji/syndiff/data/s0024/c2/k2/ffi/tess2020110032919-s0024-2-2-0180-s_ffic.fits.fz`.

Header only: DATE-OBS 2020-04-19T03:31:00.192, TSTART 1958.647330737788 BTJD, DQUALITY 0, image units electrons/s, expected 2136×2078 full-frame geometry. No array data were opened. Exact selection, inventory hash and exclusion list: [confirmation_frame_reservation_F1.json](confirmation_frame_reservation_F1.json). Reproducer: [reserve_confirmation_metadata_F1.py](reserve_confirmation_metadata_F1.py).

Before frame selection the coordinator preferred F1 over C4 for existing known computation cost: F1 source-overlap structure supports roughly 45–60 minute arms, while the large C4 overlap island can require roughly six hours per arm. This criterion does not depend on new exposure fit quality. The earlier C4 metadata-only recommendation is preserved as explicitly superseded before pixel inspection; it was never a failed or discarded science test.

Selection rule: among 1,226 cached F1 FITS files, exclude `tess2020120182919` (development image, historical faint/null/noise input and kernel reference) and `tess2020121015919` (historical second-epoch faint/noise input). Rank remaining basenames by SHA256 of fixed string `epsf-model-closure-confirmation-20261004-v1`, a newline, then basename. Choose the first with zero DQUALITY and expected units/dimensions. The first ranked exposure passed; no alternatives were inspected. A later infrastructure/input failure must be recorded, with any replacement following this same ranking and a metadata/technical failure reason rather than residual appearance.

This is **prospective new-exposure confirmation for the present locked family**, not a claim no earlier human or agent ever viewed the frame. Its shared SCC/catalogue/PS1 preparation also means it is not an independent-instrument or independent-population truth test. It adds one new independently fitted FFI, which is within Paper 1 scope; no time-series performance claim is implied.

## Minimum lock before opening confirmation outcomes

1. Exactly two families: existing A3 and A3 plus the two explicitly specified CCD colour-displacement coefficients. The model owner must freeze actual leaf names, colour definition/reference, normalization/centroid gauge and nested zero-coefficient parity. A magnitude-width or further spatial interaction is not silently added to this confirmation family.
2. Use a common prospective source catalogue, source roles, masks, pixel support, preparation version, calibration/prior policy and fixed scoring cohort. Geometry/quality eligibility may be derived from catalogue and masks; no positive-flux, S/N or residual cut is allowed to change the scored population between models.
3. Use the established physical 128-pixel tile scheme, five-way map with seed 20260929, fold 0 reserved for confirmation and the other four folds for fitting. This reuses the already verified implementation; it does not recycle the old exposure's photons. All occurrences of held physical pixels must be masked from the new training likelihood, finite-support prior terms and training initializer, as in the reviewed strict wrapper.
4. Fit this new FFI's parameters using training pixels only, with training-only photutils/WCS initialization. The old all-star A3 fit must not initialize the new confirmation model. External catalogues and a frozen common upstream preparation may be retained, with conditional scope explicit. Any new preparation/sky/kernel learning on the whole new image must be disclosed; it prevents calling the whole raw-to-output chain pixel-blind even when the subsequent ePSF fit is strict.
5. Select optimizer schedule, nuisance estimating policy, regularization and numerical tolerances from existing development evidence or the new training split only. Run identity parity, gradient/estimating-equation and rank checks without viewing confirmation residuals. If numerical quality fails, fix it on training data, without using held outcomes to pick schedules or checkpoints.
6. Freeze evaluation nuisance treatment. The current scorer re-solves source amplitudes and updates the global nuisance-prior scale on the evaluation image; if retained, the primary estimand is conditional shape quality, not absolute flux/zero-point prediction. A test of predictive amplitudes would require a different predeclared scorer.
7. Record final model and input/checkpoint hashes before one confirmation readout. A new hypothesis motivated by this held result makes it development evidence for the new hypothesis; use another prospectively reserved exposure for any expanded-family confirmation.

## Statistical multiplicity and detectable-effect statement

With one candidate already chosen on development data and now frozen, the primary new-exposure comparison is **one paired A3-versus-candidate contrast**. The earlier development search need not be retroactively counted as independent tests on this new held observation. This statement requires that no new-exposure result helped choose the candidate, domain, direction, statistic or stopping rule.

Freeze one primary paired held-shape statistic (for example the shared-domain core residual score used by the established scorer), and its spatial/shared-pixel resampling or fitted-noise reference. Report its absolute change and paired interval, not only a percentage of a tiny denominator. Overlapping star cores must not be bootstrapped as independent observations. If the interval cannot distinguish the candidate, retain the simpler baseline under the predeclared decision rule; this is not proof of equivalence without a scientifically justified equivalence margin.

Residual closure is a separate question. Predeclare the complete tested family of magnitude × actual fitted colour × detector cells and low-order image modes. Unsupported cells are excluded by fixed metadata/support gates and reported explicitly. Use one joint maximum statistic or simultaneous intervals across every residual/contrast claim included in that family; if both model residuals and their difference are used as inferential evidence, include them in the joint reference. The earlier 84 marginal features do not correct a new fully crossed grid, arbitrary radius scans, additional modes or fitted centroid controls. Exploratory plots may still be shown but labelled as such.

Report simulation exceedance counts and Monte Carlo limits, not an exact p-value inferred from zero exceedances. For orientation, zero out of 2,000 simulations has one-sided 95% Monte Carlo upper tail bound about 0.00150; 500/500 positive-control detections has a one-sided 95% lower power bound about 0.994, not certainty. These are numerical precision descriptions, not new scientific success thresholds.

Measure power on the actual new support, noise/overlap geometry and fitting procedure, for x and y shifts, width and any newly claimed colour-response mode. The existing declared amplitude grid can be reused as a sensitivity curve; do not turn 0.01 pixel or 0.3 times nominal flux error into a new scientifically accepted requirement. Include deliberate independent empirical-profile mismatches where available; same-family derivative injections measure conditional estimator sensitivity. A null generated from the fitted model tests that model's noise/estimator, not physical completeness.

Final publication can claim a supported prospective relative improvement if the locked candidate improves the predeclared score, does not trade it for material flux/centroid response failures within the independently constrained domain, and its numerical behaviour is stable. A noise-level or negligible-bias statement additionally needs calibrated power and meaningful effect upper bounds. Where no independent physical ceiling is justified, report achieved bounds and remaining uncertainty rather than invent a pass threshold.

**Conclusion:** one metadata-reserved new F1 exposure, with a locked two-family comparison and a strict held-pixel split, is the minimal practical prospective confirmation. It is materially stronger than retrospectively calling historical five-field folds blind, while preserving the single-FFI paper scope and making its conditional upstream, population and absolute-truth limits explicit.

## Frozen primary endpoint — before reserved-frame result inspection

This section replaces the example primary-statistic wording above. The primary confirmation endpoint is now fixed as

$$\Delta=\operatorname{median}_{i\in\mathcal H}\log\!\left(\frac{c_{i,19}}{c_{i,A3}}\right),\qquad c_{i,m}=\frac{1}{n_i}\sum_{p\in\mathcal P_i}\left[\frac{d_p-\widehat\mu_{p,m}}{\sigma_p}\right]^2.$$

Here $\mathcal H$ is the fixed set of eligible held stars; $\mathcal P_i$ is star $i$'s original valid radius-three-pixel core, wholly in the held physical region; $n_i$ is its fixed number of pixels (at least 15); $m$ labels either the A3 baseline or the 19-colour-coefficient candidate; $d_p$ is prepared image value at physical pixel $p$ in image flux units; $\widehat\mu_{p,m}$ is the full scene prediction after the predeclared evaluation nuisance refit in the same units; $\sigma_p$ is the common supplied image standard error; $c_{i,m}$ is the dimensionless mean squared normalized core residual; and $\Delta$ is the dimensionless median paired natural-log ratio. Negative $\Delta$ favours the candidate. Its exponent is the median paired score ratio, not an integrated flux-bias ratio.

All quantities use the same stars and physical pixels in both arms. Both scores must be finite and strictly positive; an invalid score is a recorded evaluation failure, never an arm-specific source removal or a tunable log floor. Signed source fluxes remain permitted and do not define the cohort. The endpoint is computed at the predeclared cap/stability-selected training-only checkpoint, not whichever held checkpoint scores best.

Confidence uses **paired dependence blocks**, never an individual-star bootstrap. Start from the fixed physical 128-pixel held tiles and merge units that share represented fit-input pixels or overlap-solve dependence; every resampled unit carries all its stars and both model scores. Report the resulting unit counts, largest unit, and effective dependence limits. A giant connected component cannot be converted into thousands of independent trials by counting stars or unique pixels. Preserve/recompute shared nuisance calibration in replicate scoring; alternatively state explicitly that an interval conditions on a frozen common nuisance scale and excludes that uncertainty. The existing actual scorer recalibrates its nuisance-prior scale using the evaluation contributors, so a metric-only tile bootstrap is an empirical conditional spatial-sensitivity interval unless that extra coupling is propagated. If independent units or calibrated resampling are inadequate, the numerical endpoint is still reported but its claimed inferential precision is limited/inconclusive.

A predeclared 95% paired interval below zero supports improvement of this **conditional shape endpoint**. An interval overlapping zero does not establish equivalence; the simple baseline is retained under the stated selection rule. This criterion is separate from scientific flux/centroid accuracy and residual-noise closure, which still need their own meaningful bounds.

### Fixed safety table; no additional winner search

Report the paired score ratio, source count, physical-pixel count and dependence-block count for each of these fixed marginal strata: catalogue T bins [8,9), [9,10), [10,11), [11,12), [12,13); the actual fitted colour below/above its training-only reference value; and the nine detector cells with edges 0,683,1366,2048 pixels. Also show the complete crossed occupancy/residual grid to expose unsupported intersections. Include failures, exact mask/support coverage, and fixed-denominator underlying residual images. Do not change these strata after viewing their outcomes.

Predeclared secondary structure columns are core, x/y dipole, radial/core-wing and the two quadrupole modes, using the same fixed radial window/raster rules and explicit global joint correction across the supported tested domains/model contrasts. WCS shifts, flux response and independent-profile/mismatch controls are reported alongside. These are regression/structure safety diagnostics; no subgroup becomes a replacement primary endpoint because it looks favourable. Without a justified scientific non-inferiority/bias margin, do not label a subgroup safe merely because its test is non-significant. Record upper limits and any regression requiring a scientific decision.

### Shared stages before the model branch

Sharing stages 1 and 2 is permitted because the two added coefficients are identically zero and frozen there, provided both arms use the same adapter graph, physical support, state, initialization and context. The pinned native code initializes a fresh Adam optimizer at each stage, so a stage-3 branch should start both optimizers fresh; carrying stage-2 moments would itself change the native procedure.

However, a new `scene_fit.run` invocation with `--init-from`, `--init-state-file` or `--resume` executes a fresh initial flux solve and `update_flux_prior` before the stage loop. The uninterrupted stage-2 to stage-3 transition does not automatically do this. Therefore a naïve warm restart is not exact continuation. The branch wrapper must preserve the exact saved prior/mask/source state and avoid this extra refresh, or explicitly predeclare the same refresh as a common changed schedule for both arms. Require a short zero-extra continuous-versus-branched parameter/prediction/state parity check before accepting the shared-stage optimization. Save both model metadata, reference colour, centring gauge and source/calibration lineage at the branch point.

**Frozen decision scope:** one prospectively reserved new F1 exposure, one primary paired median-log score contrast, fixed secondary safety domains and a declared conditional nuisance-refit procedure. Shared-stage computation is an efficiency measure only after context/state parity; it cannot change the scientific comparison or turn development data into blind confirmation.

### Executable paired-bootstrap specification

Freeze 10,000 resamples with seed 20261004. Begin with the 128-pixel tiles containing scored source centres; union tiles connected by any common fitted input-pixel UID or common overlap-island solve. Draw the resulting components with replacement, drawing as many components as observed, and carry **all** paired per-source log ratios in each drawn component, including multiplicities. Each replicate remains an equal-star median, not an equal-component median. Report the 2.5th and 97.5th percentiles, component count, largest component fraction, original fixed-source count and all raw paired scores. If there is only one component, report that resampling cannot estimate uncertainty; do not replace it with a star bootstrap.

This first inexpensive interval is explicitly **conditional on the learned models and each arm's realized evaluation-calibration scale**. A metric-only resample does not propagate the global scale re-estimation or all long-range detector covariance. It must therefore be labelled a conditional dependence-block interval, not complete physical uncertainty. Subsequent two-pass nuisance/noise simulations and independently calibrated shared-scale sensitivity quantify those missing terms without pretending the original metric bootstrap included them. A conditional interval upper endpoint below zero supports a conditional comparison only; final scientific acceptance retains the separate input, stability and physical-impact limits stated above.

Applying this frozen statistic to the already inspected historical C4/F1 fields is development analysis. Only the prospectively reserved new F1 exposure can provide the proposed new-exposure confirmation for the already locked family.

## Prospective source-role policy approved before new-frame outcomes

The coordinator froze the reserved-frame source roles from catalogue magnitude alone: contributors have 8 ≤ T < 11; anchors have 11 ≤ T ≤ 13; other catalogue sources are nuisance. New static/geometric core-support demotions apply in common to both families. The pre-frozen `v1min` source-ID exclusions are retained as conservative external exclusions. Old `masked_extra`, residual masks and `in_contrib_group` quality decisions are **not** copied into the fresh confirmation source-role selection.

This is an explicit common prospective population-policy change, not exact replication of the development population. If the historical `v1min` list was derived from earlier image inspection, disclose it as external historical sample selection. Its use does not inspect the reserved exposure's results, but excluded/unsupported source populations do not inherit a universal completeness or accuracy claim. Record exact list/mask hashes and counts before fitting; do not add exclusions after looking at confirmation residuals. The final T safety bin includes the upper boundary T=13 to match the frozen anchor eligibility; this resolves that endpoint explicitly for the prospective sample.

All static forward profile banks must use an analytic/header-derived or training-only seed wherever applicable. The independent native-path audit found no retained static shape bank after an explicit training-only parameter override, but the fresh FitBundle geometry/source construction must still be header/catalogue based, and any different cache-enabled path requires its own audit. This policy is approved before the selected exposure's pixel/results inspection and applies identically to A3 and CCD19.

## Prospective magnitude-family amendment — 2026-10-04 19:47:34 UTC

**This amendment supersedes the earlier two-family count and interval level.** The coordinator requested it while the reserved exposure was still in preparation/mapping, before inspecting any model or held-out result from that exposure. Reason: the original closure plan explicitly included magnitude and colour response, while the implemented new-current-input comparison had so far tested only the two CCD colour-shift coefficients. This is not a response to a reserved-frame confirmation result.

The final allowed family set is:

1. A3, the existing baseline.
2. CCD19, A3 plus the two CCD colour-displacement coefficients already specified.
3. A3w, A3 with the **historically fixed** brightness-width coefficient $b=0.8\times10^{-3}$ square TESS pixels per $10^4$ electrons per two-second read. The coefficient is not fitted or tuned.

Here $b$ is the per-axis blur-covariance coefficient in the archived `bright_width.py` convention, selected in historical `colourwidth_r2_20261001`; the code leaf is 0.8 because its unit is $10^{-3}$ square pixels per $10^4$ electrons. The native law is

$$\Delta\Sigma_i=b\frac{q_i-q_{\rm ref}}{10^4\ e^-},\qquad q_i=(2\ {\rm s})F_i P_{{\rm peak},i}.$$

Here $\Delta\Sigma_i$ is the added blur covariance per axis for source $i$ in squared TESS pixels; $q_i$ is predicted peak charge per two-second read in electrons; $q_{\rm ref}$ is the frozen training-population reference charge; $F_i$ is the consistent image-amplitude/flux predictor in electrons per second; and $P_{{\rm peak},i}$ is its rendered peak response per unit amplitude. The product $F_iP_{{\rm peak},i}$ must have the actual predicted peak image units: do not silently renormalize a finite/clipped template while retaining an amplitude defined under another normalization. The renderer applies half of $\Delta\Sigma_i$ as the raw-base-gauged Laplacian-generator weight. This first-order generator family is a predictive sensitivity, not proof of brighter-fatter detector physics.

A3w enters confirmation **only if construction gates pass before test inspection**: q units/source provenance, zero-coefficient parity on a shared renderer graph, saved q/reference/cap/source-ID roundtrip, frozen coefficient and stage labels, appropriate gradient/estimating-rule semantics, and no prohibited historical/test-information path. For the new reserved exposure, external catalogue brightness plus training-only calibration/peak-profile estimation is preferred. Do not reuse an all-star historical image-flux/q array as fresh independent truth. If native q construction cannot satisfy the declared training-only/external-data contract, the arm construction-fails without viewing its confirmation result; it is not silently repaired using the held photons. Record cap/invalid-q handling and supported domains.

No CCD19+width combination, other coefficient value, new fitted magnitude law or further family is allowed under this confirmation. A later new hypothesis requires a new prospectively reserved confirmation sample. Existing F1/F2 width development checks may establish construction/robustness before the confirmation, but may not tune this fixed b.

### Multiplicity and frozen selection rule

There are now **two** primary baseline contrasts, each using exactly the frozen paired median natural-log core-score ratio: CCD19/A3 and A3w/A3. Both use identical source/pixel support and the same dependence components. Use the **same 10,000 component-resampling draws** for both contrasts. Report two-sided **97.5% conditional intervals** for each, using the 1.25th and 98.75th bootstrap percentiles. Bonferroni then gives at least the nominal 95% family coverage for these two contrasts under the already stated conditional resampling assumptions; it does not repair missing model/global-calibration/noise uncertainty.

The selection rule is predeclared:

- If neither candidate has a conditional interval upper endpoint below zero, retain A3. This is not an equivalence or noise-closure certificate.
- If only one candidate improves, it is eligible only if the preregistered safety, conditioning and signal-preservation gates also pass.
- If both improve and pass those gates, prefer A3w: it has the same fitted parameter count as A3 and uses the historically fixed b. If A3w fails a gate but CCD19 passes, use CCD19. If neither passes, no extension is accepted.
- Do not add an unplanned candidate-versus-candidate inferential contrast or choose favourable subgroups as replacements for the primary statistic.

The residual/structure simultaneous family must likewise cover the full set of inferential claims actually reported across the three families and supported domains; the earlier 84 or 300 historical thresholds cannot be reused. The prospective fixed safety table and all existing input, convergence, nuisance-refit and physical-impact qualifications remain in force. The proposed 0.3-sigma impact budget is **not** an accepted absolute science criterion.

**Amended conclusion:** confirmation is now a frozen three-family predictive/parsimony comparison with two multiplicity-controlled baseline contrasts, conditional admission of the historically fixed-width arm through technical/provenance gates, and no further family search on the reserved result. Any selected width term remains an empirical predictive choice until independent controls support a physical interpretation.

Width q construction caution: a warm-start flux solve on the strict training image does not measure fully held sources—their amplitudes may be zero, prior-only or unconstrained. Freezing those values would silently assign the wrong held-star brightness response even if it leaked no test photons. Prefer one common catalogue-flux predictor calibrated using training pixels and multiplied by the peak of a training-only initialized profile for **all** sources. Freeze the training-population reference charge, cap, and source-ID mapping, carry that array unchanged into evaluation, and document its uncertainty. Do not mix data-derived training charges with artificially zero held charges. Any alternative conditional held-charge estimator requires an explicit prior amendment and corresponding refit/null treatment before confirmation outcomes.

The shared stage-1/2 option applies only to A3 versus CCD19 with its extra coefficients zero and frozen. A3w's fixed b is nonzero from the start, so its forward profiles and early WCS/flux updates differ even though b is not trainable. Reusing the A3 prefix is not an equivalent full A3w fit. Run the fixed-width arm's own full recipe from the common training-only initializer; delayed activation would be a different algorithm requiring an explicit prospective amendment, not an unnoticed efficiency substitution.