# Paper 1 editorial audit — 2026-10-01

## Figure layout and display revision

- Figure 1 now places (a) across the top and (b)/(c) beside each other below. Coverage is redrawn from the archived footprint and skycell-polygon coordinates; the stored PS1 assignment cutout is unchanged.
- Removed the PS1 background-component/sampling diagnostic from the active figure inventory at the user's request. Its source and exports remain recoverable on disk; the distinct full-skycell decomposition remains in §2.3.
- Figure 6b uses adjacent shared-axis panels and a common symmetric-log scale: linear within ±10 electrons/s, logarithmic outside, exactly white at zero. The symmetric outer limits follow a documented pooled-intensity rule; the data, crop and selection are unchanged.
- Section 4.4 now explicitly describes the local-polynomial penalty's noise-suppression purpose and held-out strength validation.

## Section 3.4 motivation and Section 4 data decision

- §3.4 now states the user's motivation: reliable subsequent light curves need stable scene calibration, not merely a visually clean image. Single-frame figures do not demonstrate temporal accuracy. No separate transition figure is necessary; the old in-sample spatial/colour-trend plate moves to Appendix B.
- Completed suitable development fits do exist: `prior_bakeoff_20260930/runs/F1_lp7s_full` and `F2_lp7s_full`, plus official K=5 fold fits. Their saved metadata confirms λ_local_poly = 3e8, window 7, λ_fine_nbr = 0, and reject_every = 0. The intended penalty is present; background preparation predates the new wing-mask recipe.
- Use these saved fits now for figure construction. Do not wait for final background products, silently promote development data to accepted final results, or equate the correct penalty with passing all validation gates.
- `figures/publication/section4_sources.json` selects the input root, fields, arm, folds, and required penalty settings. New builders record its hash and check the recipe; replacing the dataset also requires compatible schemas and renewed source/figure checks.
- Distinguish fitted WCS changes from measured astrometric accuracy, predicted chromatic response from observed residuals, and full-fit illustration from held-out validation. The earlier bake-off's failed non-rising-learning-curve gate remains relevant, not hidden by the figure refresh.
- Added five plates: fitted-minus-initial WCS fields (§4.2), spatial ePSF nodes (§4.3), reference and signed Gaia-XP-u response differences (§4.4), and held-out stamps plus matched five-fold scores/learning curves (§4.5). Superseded WCS, response and learning-curve plates leave the active canvas; older empirical colour/photometric diagnostics remain until equivalent replacements exist.
- Checks include the saved WCS basis and initial coefficients, pinned model source, crop-local detector nodes, pixel-integrated rendering without a second integration, and shared-source score comparisons. Stamp examples use fixed catalogue-coordinate targets and coverage, not residual quality; grey raster cells include unsampled positions.
- Recomputed P5 from individual learning-curve bins: zero of ten fold fits pass every T=8–13 bin. Worst T=10–13 end/min ratios are 1.155 (F1) and 1.123 (F2). Lower scores than the no-local-polynomial-prior control are not evidence that convergence or absolute calibration is solved.
- Verification: sixteen reproducible plates / forty-eight exports, thirty-eight active images and eighteen subsection links pass format, output/generator-hash, placement and bullet-only-outline checks. New §4 inputs match their recorded hashes. Strict historical-source validation remains failed: the dataset README and three live wing-mask inputs (`maskfoot_20261001/README.md`, `score_ksb.json`, `score_ksb.py`) changed after their older exports. Preserve the old fingerprints; those background/mask snapshots require separate reconciliation, not silent hash updates.

## Section 2–3 simplification and wing-mask revision

- Removed the unidentified supplied PS1 reference and the historical F = 1/F = 4 de-crowded pair from the active canvas at the user's request. Original assets remain on disk. Section 2.4 keeps only the template-sampling illustration.
- Retained the traceable PS1 decomposition. Its smoothed/rebinned excluded component is explicitly a counterfactual diagnostic, not background entering the processed template.
- Replaced Figure 6b's full-frame/zoom grid with two identical-coordinate, unbinned 256² cutouts and a shared linear −10 to +10 electrons/s scale. Selection uses Gaia contributor density, excluding the previously identified vertical-artifact column, not residual quality. This changes the displayed region; it does not repair the historical image.
- The replacement remains historical F1 data. Newer dataset backgrounds are provisional and were put on hold for review; their existence is not scientific acceptance.
- Updated §3.2 from obsolete fixed pad 2 to method A: magnitude-sized centred disks, harmonic fill, additional exclusions in background-fit pixels only, pad 0. Shared masks, Hotpants masks and scene selections are not enlarged by this option.
- Evidence: `maskfoot_20261001/README.md`, the timestamped 14:52–15:36 EDT dataset decisions, and `paper-pin-diff` background code. The operational linear-lane pin is a3d50a0; 846c1c3 failed its shape check and produced no accepted replacement.
- Mask sizes target a fitted wing threshold, not zero stellar light. F1 development evidence does not establish universal performance; crowded neighbours leave a wing floor, and independent cross-field background validation remains open.
- Put a compact wing-mask explanation and direct insufficient-mask comparison in §3.2. Detailed background checks belong in Appendix A; the older full-frame decomposition moves there as historical context.
- Figure 6a traces actual background-fit exclusion boundaries, including neighbouring masks, at fractional catalogue coordinates. The selected T=8.29 star has nominal radii 9/18 pixels; its nearest other 7.5≤T<13 catalogue star is 22.4 pixels away, so it is not labelled unconditionally isolated. The local annulus subtraction is a display operation, not a clean-sky truth measurement.
- Claude-history cross-check: session `37dbfffb-6438-4ed7-aced-5f5832c05295`, October 1 wing-radius discussion; executed estimator code and saved `maskfoot_20261001` masks/scores supply the plotted evidence.
- Validation: eleven reproducible plates / thirty-three exports and thirty-seven active image references; export formats, generator/output hashes, links, all eighteen subsection-menu entries and bullet-only outline checked. The existing mutable dataset-README source-hash failure remains explicit; no other input mismatch was found.

## Figure 1(c) assignment display correction

- Removed the right-hand skycell overview and the four-pixel-only highlighting. The same 211 × 207 PS1 cutout now colours every native TESS destination separately, including partial pixels at its edges.
- Direct map check: 43,677 PS1 pixels, 16 native TESS destinations, zero stored −1 entries. The earlier “other TESS pixels” class was a plotting simplification, not unmapped data; the unconditional “unmapped” legend entry was unnecessary here.
- Stored −1 is local to a skycell's index map. Pixels outside the mapped region or owned through overlapping skycells can remain unassigned there; this value alone does not demonstrate a hole in the combined TESS coverage. No production mapping was changed.

## Active-canvas cleanup — user-directed follow-up

- Removed the separate previous-figure desk and its navigation links. Latest available figures now appear directly beside the relevant outline text or in the appropriate appendix; none is hidden in an archive dropdown.
- Kept 28 distinct existing figures, including replication and supporting diagnostics that have no replacement. Historical-data warnings remain explicit; “latest available” does not mean final-recipe validation.
- Omitted six replaced assets: old F1 coverage, F5 workflow, incorrect F5 rendering inset, F17 in-sample preview, F20 kernel mosaic and G6 refit mosaic. The expanded/corrected publication plates supply their active replacements; provisional held-out F17 remains visible.
- Each active image appears once. All 34 distinct images in the former desk have an explicit keep/replace decision checked by the canvas validator. Original files remain recoverable; this change removes obsolete displays, not scientific data.

## Section 2–3 revision — user-directed follow-up

- The user retained the existing narrative and requested more concrete method illustrations; the reorganization below remains a proposal, not the adopted outline.
- §2.1 now distinguishes candidate-skycell matching, overlap ownership, and the stored PS1-to-TESS index map. The code chooses the closest projection centre among overlapping candidates, not necessarily the closest skycell centre.
- §2.3 now explains why PS1 background discontinuities must not enter the contaminating-source template; this is separate from the exposure-specific TESS background in §3.2.
- Historical/diagnostic examples are explicitly labelled in the active sections rather than left only in the archive. They do not close the final-data or controlled-sampling acceptance gates.
- The supplied three-panel chip-gap example is preserved unchanged but remains unidentified: searches of the legacy SynDiff source/notebooks did not recover its generating code or skycell. It is a motivation reference, not publication evidence. The new traceable cutout diagnostic must not be described as reproducing that gap or measuring a cross-gap offset.
- Added §3.3, preliminary-image comparison; the former §3.3 “Why fit a scene?” is now §3.4. Sections 4 onward retain their structure.
- Figure review: new mapping panel (c) reads stored skycell 2483.062 assignments, groups F = 4 indices by native pixel, and distinguishes unmapped pixels. The preliminary comparison uses the exact historical FFI-minus-background/Hotpants pair, one shared signed asinh display, and the same zoom; 27 contributor markers are not a source-retention truth catalog. The historical blur's unquantified seam-cell error remains explicit.
- The traceable PS1 example uses skycell 2436.091 and a segmentation-defined excluded component, not a fitted smooth sky. Gaussian smoothing and nominal area means are a controlled illustration, not actual WCS-mapped TESS flux or a recovered chip-gap experiment.
- Separate provenance files preserve the original seven-plate snapshot and its known mutable-README hash drift; adding figures does not retroactively validate that snapshot.
- Verification: all ten plates / thirty exports pass size, resolution, embedded-font, SVG, generator/output-hash, local-link and placement checks; the menu matches all eighteen subsections and the outline remains bullet-only. All new input fingerprints match. Strict validation still fails only on the previously recorded live dataset README drift. No browser screenshot or production rerun was performed.

## Structure and figure review — 15:15 UTC snapshot

- Verdict: strong methodological material; not yet a submission-ready evidence sequence. Simplify the narrative before adding more diagnostics.
- Scope: current outline, old figure plan/canvas, generator/manifest and updated dataset record; visually reviewed all 34 distinct historical PNGs and seven new plates. The old canvas repeats one image, giving 35 historical placements.
- This is a structural, visual and provenance audit, not an independent rerun of every historical measurement or a completed literature audit. Contact-sheet review does not replace final-size proofing.
- Implemented: subsection navigation and figures beside their relevant subsections. Manuscript reorganization and renumbering below are recommendations, not silently applied decisions.
- Priorities: sample/fold diagram; held-out astrometric and pixel-prediction results; matched final-image comparison; five-FFI summary. More kernel mosaics will not fill those gaps.
- Recommend eight core main-text figures, a ninth only if targeted extraction is validated, and two tables: data/recipe and per-FFI performance. This is an editorial recommendation, not a journal limit.

## Structural findings

- High priority: methods and results are interleaved. Current §4.2/4.4 describe the model but also request performance figures; §5 mixes kernel construction, image evidence, stability and extraction. Separate computation from demonstration.
- High priority: introduce data/evaluation before methods: exposure IDs, epochs, sample roles, masks, colour coverage, folds and scoring rules. Five FFIs spanning four cameras do not establish universal performance.
- High priority: define accuracy operationally. Explain held-out position measurement with shared calibration fixed, scoring pixels, permitted nuisance-flux fits and Gaia/XP uncertainty. A WCS correction map is not an accuracy map.
- High priority: separate calibration, de-crowding and extraction endpoints. Refit stability is neither external accuracy nor proof that targeted photometry works.
- Medium priority: template engineering dominates the opening. Keep the resolved-scene illustration; move background iterations, sampling controls and seam diagnostics to appendices.
- Medium priority: F7 repeats F17; F10/F12/F15 overlap; F14 plus its initialization grid is excessive. Give each main figure one question; distinguish model fields from empirical measurements.
- Medium priority: numbering reflects development history. F20 precedes F18; the F5 inset sits in another section; F8/F13 lack active slots. Keep asset IDs now; renumber after settling the figure list.
- Medium priority: older `FIGURE_PLAN.md` gives F6/F7 different meanings and describes superseded background processing. Treat it as historical, not a second governing outline.

## Recommended manuscript and figure placement

| Proposed location | Purpose and figure | Assets / action |
| --- | --- | --- |
| 1. Introduction | Question, contribution, testable endpoints | No results gallery; colour dependence itself is not the novelty. |
| 2. Data and evaluation | M1: coverage, star roles/folds, colour/magnitude coverage | F1 + new sample/fold panel. Table 1: five exposures, epochs and per-stage recipe pins. |
| 3.1 Method overview | M2: end-to-end flow | New F5; add explicit TESS input and separate held-out scoring branch. |
| 3.2 Reference scene | M3: resolved scene, removed segments, retained neighbours | F3 + common-coordinate segmentation zoom; F2 bandpasses and full decomposition in appendices. |
| 3.3 Preliminary image | Bootstrap and image-level background | Compact text; final harmonic F6 and blank-region validation in Appendix A. |
| 3.4 Joint calibration | M4: spatial response example and rendering | Corrected F5 inset + small final F14 profile panel; full mosaic/gauges in Appendix B. |
| 3.5 Per-band subtraction | Remapping, response-to-kernel conversion and scale | F20 basis/reconstruction in Appendix B; no performance claims here. |
| 4.1 Chromatic pixel response | M5: empirical residual stacks + displacement trend | Selected F10/F15 + compact held-out F11 data/model/residual examples; full F12 fields in Appendix B. |
| 4.2 Astrometric validation | M6: held-out vector residuals + radial-error distribution | Rebuild F9; common stars/masks, bias and scatter separately. |
| 4.3 Photometric consistency | M7: held-out colour/magnitude trends + spatial residuals | Rebuild F17; F7 stays development history, not a duplicate main figure. |
| 4.4 De-crowding performance | M8: image comparison + population residual score | F18 + selected F16 metrics/stamps; separate full-chain and kernel-only effects. Table 2: all five FFIs. |
| 4.5 Extraction, conditional | M9: target retained/subtracted + flux/uncertainty checks | Merge F19/F21 only with real independent targets; otherwise future work. |
| 5. Discussion; 6. Conclusions | Limits, interpretation, reproducibility, supported findings | G6 detail in Appendix B; failures remain explicit in the discussion. |

- Appendix A: sampling, seams, flux accounting, harmonic background and masks.
- Appendix B: full ePSF/spatial fields, gauges, priors, convergence, ablations, kernel reconstruction and refit.
- Appendix C: bandpass conventions, exact D13/D18 weights, spectral holdouts and outliers.
- Promote an appendix diagnostic if it changes the conclusion; do not hide failed gates in supplementary material.

## Individual main-figure asset audit

| Working ID | Visual/scientific verdict | Required action / destination |
| --- | --- | --- |
| [F1](../f01.png) | Clear overview; the 986-skycell grid shows mapping, not measured valid coverage. | Verify final mapping/count; add usable coverage and roles/folds; M1. |
| [F2](../f02.png) | Readable curves; a synthesized response curve does not establish accurate synthetic fluxes. | State normalization, photon/energy and zero-point conventions; exact weights; Appendix C with spectral residuals. |
| [F3](../f03.png) | Good accounting concept; whole-skycell view cannot establish neighbour preservation. Black conflates low/invalid pixels. | Adopted-recipe segmentation zoom; distinguish invalid pixels; M3, full plate appendix. |
| [F4](../f04.png) | Labelled nominal-grid demonstration; sharper fine pixels are not a measured accuracy gain. | Illustration or fixed-WCS sampling experiment; Appendix A. |
| [F5 workflow](f05_workflow.png) | New layout is readable; TESS input is implicit, validation only a footnote, refit/extraction conflated. | Explicit TESS/Gaia inputs and scoring branch; distinguish optional extraction; M2. Old panoramic F5 is too small at print width. |
| [F5 rendering](f05_rendering.png) | Corrects the old sixteen-sample diagram; symbols not all defined in caption. | Define coordinates, grid centre, stored response and normalization; distinguish spatial-node from pixel-grid interpolation; M4. Never reuse old `f05inset.png`. |
| [F6](f06_background.png) | Signed difference is useful; nearly identical cleaned full frames add little; legacy background recipe. | Harmonic output, mask, local blank-region residuals and validation; Appendix A. |
| [F7](f07_trends.png) | Spatial maps flatten but A3 colour trend remains strong. Jointly clipped sample; separate zero points. | No blanket photometric-gain or absolute-calibration claim. Merge concept into held-out F17; archive this comparison. |
| [F8](../f08.png) | Synthetic reconstruction, not observed performance. Bottom text is cramped; displayed F4 non-core RMS exceeds F1 here. | Not evidence of F4 superiority. Redesign controlled experiment/layout; Appendix A. |
| [F9](../f09.png) | Fitted-minus-production WCS measures solution change, not accuracy; component scales differ. | Held-out residual maps/CDF for M6; correction map only supporting context, with vector key. |
| F10 / F10F2 | Strong empirical-image layout: blue/red/difference, baseline/adopted, shuffle null. Dense lower mosaics; old #14 called adopted inside figure. | Final A3 heldouts, matched magnitude/crowding/phase, counts and uncertainty; selected stacks + F15 trend for M5, replication appendix. |
| F11, missing current evidence | Needed to show actual observed/model/residual pixels, not only summaries. | Preselected isolated/crowded and red/blue cases; same scale, masks, coordinates and IDs; compact M5 subset. |
| [F12](../f12.png) | Useful spatial response, but smooth model fields can be mistaken for independent empirical estimates. | Model-only labels, calibrated colour/radius domain, empirical checks; A3 rebuild in Appendix B. |
| F13, optional | No active slot; model selection already has supporting diagnostics. | Do not add a redundant main plate; held-out ablation table/plot in Appendix B. |
| F14 / F14b | Tiled-coordinate axes obscure detector location; response units unspecified. The 72-stamp initial/final grid is too dense. | Small spatially located final examples/profiles with shared scale/support for M4; full node grid appendix; remove run-name titles. |
| F15 / F15F2 | Good radial/tangential null comparison; nine panels, role split and secondary axes dilute it. Baseline/corrected ranges differ. | Compact M5 trend; label zoomed range or add shared-scale context; role/null/replication appendix. |
| F16a / F16b | Stacks support the metrics; score plot has overlapping long titles, method IDs and overplotted points. Dry-run evidence. | Paired summaries, uncertainty/counts, common masks/noise, readable labels and matching stamps; M8. |
| F17 / F17pre | Older provisional held-out evidence exists, but not for final recipe; twelve panels and aperture decomposition overload it. | Final common-fold trends/map in M7; aperture decomposition appendix. Train zero points off test data or scope score explicitly to zero-point-removed consistency. |
| [F18](f18_cutouts.png) | Honest fixed region/shared signed scale; retained-star saturation dominates, masks are busy, differences hard to see. | Context panel, retained-star markers, predetermined smaller zooms and paired differences; saturation/valid-pixel fractions; M8. No favourable-residual selection. |
| F19, missing | Target retention is distinct from a generic final-image comparison. | IDs, neighbours, unchanged controls and segment accounting; conditional M9. |
| [F20](f20_kernels.png) | Clear P-versus-K distinction. One node and ±4-pixel view cannot establish full-support ringing/reconstruction. | Explicit colour-bar units/normalization, full-support checks and reconstruction residuals at centre/edge; Appendix B. |
| F21, missing | No current product establishes extraction recovery/uncertainty calibration. | Real independent targets and scoped injection/recovery or repeats; conditional M9; no temporal-precision claim. |
| [G6](g06_refit.png) | Width failure is clear; flux medians hide tails; connecting RMS and maximum implies a continuous axis. | Separate categorical markers, tail/quantile checks and parameter maps; repeat final recipe; Appendix B plus limitations. |

## Historical supporting figures — disposition

| Assets | Finding / action |
| --- | --- |
| `fA_f1`, `fA_f4` | Same region/scale, but WCS/recipe differ; prominent vertical seam. Matched experiment needed; Appendix A/history. |
| `fS_lcurve` | Important generalization warning; five thin panels and dense legend. Rebuild final-recipe curves in 2 × 3 layout; Appendix B. Old curves do not set new stopping point. |
| `fA1` | Model selection, optical-axis scan and covariance answer three questions; tiny annotations. Split/trim; selection set is not a final test set. |
| `fA2` | Useful ablation/checkpoint drift, but older model called adopted. Final matched held-out ablation; separate convergence panel. |
| `fA3` | Relevant gauge/normalization sensitivity; too many tests. Split normalization from gauge/loss; show underlying residual images. |
| `fB4` | Readable A3 development comparison, warm-started and in-sample. Historical rationale, not unbiased confirmation. |
| `fB5` | Eleven-FFI development breadth is not the five-FFI adopted dataset; eight metrics are dense. Do not substitute for final Table 2. |
| `fB6` | Heterogeneous coefficient units, colours-only detector key, little physical interpretation. Parameter table with uncertainty/induced response, not main text. |
| `fB7` | Spatial width discrepancy remains visible: valuable negative evidence. Rebuild with final recipe, counts and residual stamps; discuss persistent failure. |
| old `f20`, `fG6` | Too many tiny kernel/refit panels. Compact new plates are better starting points; full grids remain diagnostics. |
| `fS_bandpass` | Useful spectral-type/colour residuals; crowded legend and extrapolated tails. Identify training/independent spectra and supported domain; Appendix C. |
| `fS_response` | Eight panels mix image residuals and catalog photometry. Split by question; response attribution remains a hypothesis without controlled evidence. |
| `fS_weights` | Lower score coexists with aperture/colour discrepancy. Paired controls, uncertainty definitions and failure discussion; no blanket gain. |

## Visual evidence behind the main cautions

- F7: spatial and colour diagnostics do not show the same improvement; development evidence, not a new measurement.

![Development spatial and colour trends](f07_trends.png)

- F18: saturated retained stars dominate; keep context but add smaller predetermined comparisons.

![Development real-image comparison](f18_cutouts.png)

- G6: positional/median-flux passes do not erase the effective-width failure.

![Development refit checks](g06_refit.png)

## Current provenance findings

- Live README now records base pin `648c0bcf4b559691177f7b4764f721ac54aebdc8` and chain pin `e795e8e16ee8ed598e629cdfe950a7ad6a10b5f0`; earlier mapping SHA `c466fb2` is documented as mapping-code equivalent. Record stage × field provenance, not a fictitious single executed SHA.
- Recorded configuration: provisional `v1min` exclusions, not `v1`; harmonic background with pad 2 in background estimation/kernel fitting. Later pin table/log supersede the README's older blocking/planned sections.
- Source reports completed os4 processing/spot checks for F2/C1/C4 and pending downstream work. Presence verification is not seam/science/held-out acceptance; this audit did not inspect all live jobs or stage outputs.
- One of 14 figure-input hashes drifted: dataset README used by F5 workflow. Stored: `6955ee77eed5b4f25e13cdf66c75b3aceed3f82baac572639e14766941cf9481`; current at audit: `30926d499f4925c11bd192c559c39ef2090b53519c14bb67610bba8721778e10`.
- Strict reproducibility validation fails source identity, although saved exports are intact. Preserve old hashes; do not silently re-stamp. Next rebuild should pin immutable recipe snapshots and compare against execution manifests.
- Only seven new plates have the new export manifests; three displayed legacy plates and historical assets are not covered. Extend provenance when promoting each figure.
- Manifest `export_status` still requests visual review despite the prior screen review. Next release needs explicit review date/scope, distinct from final-size proofing.
- Figure generation was not changed/re-run for this audit. No fits, pipeline stages, production configuration or historical arrays were modified.

## Navigation and audit checks

- All 17 subsection links match the source outline and resolve to visible headings, including figures-only mode; all seven new figures sit in their assigned subsections.
- Canvas/audit local links, three audit tables/images, export hashes/formats and bullet-only outline checks pass. JavaScript smoke tests pass for all three view buttons and deep-link highlighting.
- Strict source validation still fails on the one README drift above. Browser layout was not screenshot-tested from this remote task; refresh the existing preview to see the navigation update.

## Prioritized completion sequence

- First: settle eight-core-figure story and freeze data/evaluation definitions; sample/fold diagram before scoring.
- Next: final held-out pixel/chromatic, astrometric and photometric evidence, including failures/counts/uncertainty definitions.
- Then: matched image controls and cross-FFI table; seam, weight, scale and reconstruction acceptance.
- Last: proof selected plates at print size; remove run-name titles, standardize legends/units, check colour/grayscale readability.
- Conditional: extraction only with its own evidence; do not expand scope merely to fill F19/F21.
- Conclusion: viable method story, but publish-ready figures need a smaller hierarchy and final controlled evidence. Polished development panels do not replace those results.

## Version control and authority

- Paper source lives in the separate `dev` repository, not the parent pipeline repository.
- Editing branch: `codex/paper1-publication`; base commit `8ee5d57`.
- Worktree: `dev/.worktrees/paper1-publication`; original checkout and canvas remain unchanged.
- Two existing uncommitted rendering-inset exports were copied into this worktree before editing; they are preserved and are not the corrected new inset.
- `OUTLINE.md` is the current editorial outline. The prior outline is preserved in `archive/OUTLINE_20260930.md` and Git.
- Recipe authority for this snapshot: `/astro/armin/koji/syndiff/dev_runs/paper_dataset_20261001/README.md`, read 2026-10-01. It supersedes older outline and freeze-log entries, not necessarily later owner updates.
- The initial checkpoint preceded the code pin; the current snapshot and exclusions are corrected above. This audit is not a live monitor.

## Corrections applied

- Rendering: pixel-integrated E needs one bilinear sample times O², not a sum of sixteen evaluations. Checked `forward_epsf_wcs/epsf_model.py`, `render_stamps` and `render_physical_pixels_blocksum`; the latter retains a legacy function name.
- Priors: adopted lp7s local-polynomial recipe replaces moment-blind; brightness-width and fitted background are excluded from the five-FFI dataset.
- Initialization: uniform header-WCS mapping for the five-FFI dataset, not the older temporal-WCS initialization.
- Scope: five independently fitted FFIs cover four cameras, not only three development fields.
- Full-chain comparison: remapping and fitted scale also change. Only matched controls isolate kernel effects.
- Kernel model: P₀/P₁ are a normalized least-squares intercept/slope through 25 colour quantiles, not an exact response/derivative pair at one reference colour. Checked `e2e_f1_20260930/scripts/kernels/k02_band_epsf.py`.
- Matching kernels K differ from target responses E; source-position spatial weights precede convolution.
- Separate D13 synthesis weights from D18 fixed colour labels; old kernel metadata legitimately records older colour-label weights, not necessarily adopted synthesis weights.
- Refit: effective position and binned flux satisfy the older D14 tolerances, but effective width fails. Do not report a blanket stability pass.
- Accuracy: in-sample previews and model-selected/clipped samples are not held-out performance evidence.
- Catalog flux, parallax implementation, universal sub-percent synthesis, sampling superiority, and response-curve explanations require evidence; unsupported assertions removed or qualified.

## New figure exports

- Seven plates: corrected workflow and rendering inset (F5), background decomposition (F6), spatial/colour trends (F7), matched real-image cutouts (F18), response/kernel bases (F20), refit diagnostic (G6).
- Formats: 300-dpi PNG, PDF with TrueType font setting, editable-text SVG; 7.1-inch page-width layouts.
- All data panels use existing saved products; no invented observations, interpolation of missing experiments, or rerun fits.
- Captions identify development recipes, inherited selection, units, scales, and limitations. No development plate is labelled final science.
- Each build records source/output SHA-256 hashes, generator hash, display settings, and relevant numerical tables in `figures/report_20260929/publication/manifest.json`.
- Full-image F6 display uses 8×8 area means only; original FITS are unchanged.
- F7 inherits a common clipped in-sample table; no new claim of unbiased gain is permitted.
- F18 uses the fixed central 256² region with the union of method masks outlined. It is not a hand-picked best residual; all finite pixels remain visible.
- F20 uses a common ±4-pixel view with separate signed scales; normalization is over full stored supports, not the displayed crop. It is not a completed colour-extrapolation validation.
- G6 plots recorded bin medians and effective-width values; no uncertainty estimates are invented.

## Remaining figure gates

- F1: reconcile footprint/skycell count with final header-WCS mapping and valid coverage.
- F2 / Appendix C: reconcile exact response/weight metadata and held-out spectral table; older numerical accuracy claims are not repeated here.
- F3: regenerate full decomposition using adopted weights and seam fix; add neighbour-preservation segmentation zoom.
- F4 / Appendix A: matched WCS and scene for a controlled F = 1/F = 4 comparison.
- F5 workflow: checked against current dataset recipe; reconcile with the eventual pinned execution manifest.
- F6: replace legacy background example with adopted harmonic products and withheld-region validation.
- F7/F9/F10/F11/F12/F14/F15/F17: final A3 fits, frozen folds, leakage checks, and scoring tables. Older #14 and in-sample panels remain history.
- F16/F18/F20: final per-band products with matching masks/noise, seam and reconstruction gates, and known scale conventions.
- F19/F21: real targeted extraction products and independently selected targets; no current substitute.
- G6: repeat all stability gates with the adopted no-brightness/no-background recipe.
- Cross-camera summary: add after all five FFIs pass the same pipeline and validation gates; no mixed-recipe headline table.

## Submission acceptance checks

- One pinned code SHA and per-frame manifest; exact input/exclusion/fold hashes.
- Same recipe, masks, calibration scope, and metric definitions across methods/cameras.
- No training/holdout leakage, including blends and preprocessing; uncertainty/scatter definitions stated.
- Figures inspected at intended print size; labels legible; no panel/colour-bar clipping.
- Axes, units, signed scales, mask conventions, N, selection, support, and normalization documented.
- PDF vector text/lines preserved where appropriate; raster panels at adequate resolution; no manuscript titles baked into figures.
- Scientific captions agree with displayed data and manifest; all claims trace to primary outputs.
- Bibliography keys and primary literature still need a citation-by-citation audit; existing literature notes alone are not external verification.

## Checks completed for the previous figure checkpoint

- Visually inspected all seven new PNG plates; adjusted rendering-equation spacing, exposed the small background difference on an explicitly separate scale, retained masked-region image context, and used a common kernel-core view.
- Read-only validator passed: 7 plates / 21 exports; 300-dpi PNG geometry; embedded TrueType PDF fonts with no Type 3 fonts; parseable SVG; 14 source hashes; output hashes; all canvas links and unique anchors; synchronized bullet-only outline.
- Canvas contains 45 image references, including all 35 panels from the existing desk, plus new/current illustrations. Existing history is labelled rather than silently reclassified.
- Scientific readiness remains separate from export quality. Development data, incomplete holdouts, and unresolved gates prevent declaring the full figure set submission-ready.
- That checkpoint's all-source-hashes pass no longer holds for the mutable dataset README; see current provenance findings above.
- The remote Mac/plscience4 preview connection was not independently observable from this task; shared-path/link checks passed and browser navigation was requested.

## Rebuild and preview

- In the `syndiff` environment, from this worktree: `python paper/paper1_single_ffi/figures/publication/build_figures.py`, then `python paper/paper1_single_ffi/figures/publication/build_canvas.py`.
- New canvas: `figures/report_20260929/publication/index.html`; source bullets come directly from `OUTLINE.md`.
- The original preview host is plscience4. This task cannot SSH to it; a shared-filesystem preview link avoids changing its server process.
- No production jobs, large data copies, paid compute, upstream code changes, or external publication were performed.