# Independent C4 ledger visual and generation review — 2026-10-04

**Conclusion:** the completed ledger accounting is numerically reproduced and visually consistent for the **original C4 field/os4 bootstrap template**, hash `20e928283bb334f311486c4767c9f31c4b9c4c77e7d59d3b73bfa93775eb5783`. This supports captured-deletion transport and historical published-template reconstruction for that generation. It does **not** certify the latest `template_v3` field/os4 template, hash `fc400b873f5c5f40a21f24604ba02acca321219180c37792ddf47ef305ecf473`, whose decoded image genuinely differs. It also does not certify catalogue completeness, correct astrophysical attribution of blended/PS1-only detections, complete PSF removal, ideal seam physics, or unbiased photometry.

This is an independent review record. Owner summaries, flags, trigger markers, source, jobs and science artifacts were not modified. The owner summary remains `numerically_validated_pending_visual_review`; this document supplies independent scoped visual evidence without changing the owner's workflow state.

## Exact generation and method

Authoritative ledger inputs are `/astro/armin/koji/syndiff/dev_runs/ps1_ledger_fix_20261002/inventory/C4.json`, section `original_inputs`. Its template path is:

`/astro/armin/koji/syndiff/dev_runs/paper_dataset_20261001/C4/bootstrap/templates/oversampling_4/fits/syndiff_field_s0023_4_2_os4_gid0.fits.fz`

The newer comparator is:

`/astro/armin/koji/syndiff/dev_runs/paper_dataset_20261001/template_v3/C4/bootstrap/templates/oversampling_4/fits/syndiff_field_s0023_4_2_os4_gid0.fits.fz`

Both arrays have shape 8256×8256. Both file SHA256 values were independently recomputed, and the saved ledger `after` array was compared against decoded FITS data, not compressed bytes. It equals the original template exactly. Against the newer template, 2,653,149 of 68,161,536 finite pixels differ; maximum absolute difference is 1207.990234375 and full-image RMS difference is 0.6120387234 in native template units. These are real pixel differences, not gzip metadata or FITS-header differences. The underlying absolute and difference images are shown below.

![Original and newer templates and actual difference images](ledger_c4_generation_comparison.png)

The executed ledger pin is `/astro/armin/koji/syndiff/dev_runs/ps1_ledger_fix_20261002/code/pin_d3549ee`, revision `d3549ee83c0d58b89d94ceb7545d46d4140877b5`. Its `field_validation.py` resolves the recorded v2 convolved inputs, historical renderer, captured explicit deletion images, frozen rim/ownership assignments, and real sparse L5 binning. `reduce_field` adds recipient contributions in float64, checks reconstructed `after.astype(float32)` against the original template, and tests the paired accounting identity before storing float32 display arrays. Historical input renderer provenance remains distinct from latest v3.

The identity reviewed is

$$R_p=B_p-A_p-D_p.$$

Here $p$ indexes an oversampled template pixel; $B_p$ is the counterfactual before explicit captured deletion, $A_p$ is the reconstructed published after-deletion image, $D_p$ is transported captured deleted signal, and $R_p$ is the closure residual. All four are in the native template flux units. The stored residual was computed from float64 accumulation; recomputing it from the three separately rounded float32 display arrays need not give the same tiny residual.

## Full-field accounting and seams

The completed summary is `/astro/armin/koji/syndiff/dev_runs/ps1_ledger_fix_20261002/field_validation/C4/summary.json`; accounting arrays are `template_accounting.npz` beside it. Independently read all 975 recipient `report.json` files:

| Check | Result |
|---|---:|
| Validated recipient reports | 975 / 975 |
| Exact historical published contribution reconstruction | 975 / 975 |
| Reported transport and L5 arithmetic closure | 975 / 975 |
| Captured inter-skycell patches | 5,014 |
| Missing inter-skycell patches | 0 |
| Recipients with cross-projection donor names | 157 |
| Maximum saved full-field closure residual | 0.001661628485 |
| Integrated transported deleted signal | 1,240,935,607.7526164 |

The integrated quantity is a sum over the oversampled template accounting grid in native units; it is not asserted to be calibrated total stellar flux or a native-TESS-pixel aperture sum. Summing the independently loaded saved float32 deletion image in float64 gives 1,240,935,607.7526157, consistent at rounding precision.

Full-field panels use shared scales for before, after and deletion. The closure panel uses a much narrower ±0.002 scale. A full-field view averages 8×8 oversampled pixels for display only, so the stress patches below retain native oversampled pixels to avoid hiding local peaks.

![Full-field absolute, deletion and closure images](ledger_c4_full_accounting.png)

Stress patches were selected deterministically at the largest deletion, largest saved closure residual, a cross-projection recipient (`skycell.2332.095`) near the padded image edge, and the recipient with largest reported L5 closure residual (`skycell.2441.065`). The two largest-closure selections are nearly the same location, explicitly retained rather than mistaken for independent cases. Selected recipient contribution arrays were also independently checked against their saved arithmetic bounds; detailed ratios and exact report copies are in `ledger_review_evidence.json`.

![Deletion, cross-projection/edge and worst-closure patches](ledger_c4_stress_patches.png)

Visual assessment: the deletion image follows the removed compact and extended footprints; the published after image retains substantial surrounding light, including an obvious bright halo around the largest deletion near the edge. The worst closure features are small localized core/rounding structures on the ±0.002 scale, rather than a gross missing source or continuous seam. The cross-projection/edge patch is consistent with the accounting identity, with no gross accounting discontinuity visible on the displayed scales. These views support arithmetic and historical-operator reproduction; they cannot establish that the original seam algorithm or removal prescription was astrophysically ideal.

The original pilot figure `/astro/armin/koji/syndiff/dev_runs/ps1_ledger_fix_20261002/figs/seam_closure.png` was inspected too. Its title explicitly says one-cell intervention, neighbours fixed, not whole-template certification. It is supporting historical operator evidence, not a substitute for the full-field result.

## Source-loss interpretation and existing pilot images

Inspected `/astro/armin/koji/syndiff/dev_runs/ps1_ledger_fix_20261002/figs/ps1_only_examples.png`. Four examples show original PS1, pre-removal support, published post-removal image, and explicit deletion, with Gaia/PS1 positional markers. Substantial faint-source/halo flux is visibly erased inside bright-source deletion footprints. The labels correctly retain “no confirmed Gaia ID” and do not automatically promote PS1 detections to certified individual stars. These images support captured pixel-loss accounting, not an assertion that each candidate has an unambiguous astrophysical identity or independently measured deleted stellar flux.

![Previously generated source-loss examples, inspected in this review](../../ps1_ledger_fix_20261002/figs/ps1_only_examples.png)

The owner README and watcher follow-up still describe the earlier memory hold and incomplete reduction; those notices are stale relative to the newly present numerical summary. This review does not rewrite another owner's history or status. The parent campaign should link this independent record while preserving generation and scope.

## Reproduction and inspected artifacts

Own code: `review_ledger.py` and `ledger_generation_plot.py` in this directory; activate `syndiff` first. These read archived artifacts and save only derivative figures/evidence under the closure `noise/` directory. Machine-readable evidence: `ledger_review_evidence.json`; execution transcript: `ledger_review.log`. All three new figures and both original pilot figures named above were opened and inspected before this review.

**Final conclusion:** accept the limited claim that captured explicit deletions are accounted for through the frozen original C4 ownership/remap/binning operator and reproduce its published template. Keep the current latest-template ledger gate open: its decoded template is different and requires its own provenance-matched transport/accounting validation. No result here establishes complete source removal or photometric bias limits.

### Generation-comparison image inspection

The newer-minus-original full-field image shows many narrow curved diagonal stripes. Its largest-difference patch shows signed stellar features clipped by a sharp diagonal boundary. This morphology is consistent with a seam/ownership operator change and directly demonstrates why old-generation accounting cannot be silently transferred to the latest template. It does not by itself determine which generation is scientifically better.

**Conclusion remains:** historical original-template accounting passes this scoped independent review; latest-template accounting and science-quality certification remain separate, open checks.