> **2026-10-04 follow-up:** [NEIGHBOUR_SKY.md](NEIGHBOUR_SKY.md) tests an actual sky-normalization correction. It passes the original profile family but fails a broader independently fixed family; persistent real group trends remain. No new correction is adopted.

# Source-consistent signals expose sky-normalization leakage — 2026-10-04

**Conclusion: the current candidate fails the stronger signal-preservation
test.** Adding each known source to every affected cutout raises the worst
primary amplitude error from 0.48120% to **1.55152%**, above the unchanged
0.5% criterion. A controlled derivative decomposition locates the main error
in neighbour light admitted to the sky ring, which changes the aperture flux
used to normalize the learned profile. No source is removed to pass, and no
new correction is adopted. The full physical/photometric goal remains unmet.

Run: `/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/`.
Code: `/home/kshukawa/syndiff_pipeline/dev/.worktrees/temporal-raw-epsf/`,
local branch `codex/temporal-raw-epsf`; scripts `temporal_raw_epsf/source_consistent_*.py`.
This follows [the neighbour/overlap finding](NEIGHBOUR_LEAKAGE_20261003.md).
The [existing investigation canvas](http://localhost:18888/temporal-raw-epsf/canvas/temporal_raw_epsf.html)
contains this milestone and the complete preceding sequence.

## One source on one detector grid

The old primary injection modified each star's own science stamp only. The
new experiment gives each of all 172 science sources one absolute detector
trajectory and one analytic profile, then inserts it into all affected original
431 cutouts on all 1,224 quality-zero exposures. The same 235 references,
165 eligible science stars and all-172 amplitude criterion remain.

The added pixel rate is

$$\delta I_{t,p}=\sum_s a_{s,t}P^{\rm truth}_{s,t,p},\qquad
a_{s,t}=\frac{u_{s,t}A^{\rm raw}_{s,t}}
{\sum_p w^{(5)}_{s,t,p}P^{\rm truth}_{s,t,p}}.$$

Here $t$ indexes an exposure, $p$ an absolute detector pixel and $s$ an
injected source; $\delta I$ and $a$ are in electrons/second. The dimensionless
$P^{\rm truth}$ is the same independent, pixel-integrated asymmetric Gaussian
mixture used by the preceding local primary test, with unit whole-plane
integral. $u$ is the original dimensionless 0.001-amplitude sine, with its
original source-dependent phase and 5/10/20/30-day period. $A^{\rm raw}$ is
that source's recorded radius-5 aperture rate in electrons/second, and
$w^{(5)}$ is its existing dimensionless response-aware aperture/sky operator.
Thus each source's own aperture addition preserves the old signal definition.

A source's absolute position is its recorded injection position plus its own
cutout origin. Each receiving cutout subtracts its own origin from that SAME
absolute trajectory. There is no target-dependent position anchor, profile,
or injected source flux. Connected overlapping cutouts receive the same ordered
source set; their shared pixel additions are bitwise identical.

No catalogue distance or brightness cut determines support. Conservative
positive Gaussian-mixture box integrals over the complete motion envelope
bound omitted absolute pixel charge and linear r3/r5 aperture effects.
The declared limit is 1e-6 ppm; the largest resulting aperture bound is
1.05e-9 ppm. This is not a bound on every nonlinear downstream operation or
on real, unknown PSF wings. There are 189 included source/cutout links:
172 own-source links and 17 cross-source links. Two science cutouts and
15 reference cutouts receive cross-source light in this particular analytic
family. All 14 overlapping cutout pairs have bitwise identical additions.

Four focused tests pass: independent analytic profile/derivative checks,
cutout translation, conservative tail bounds, and connected support. Own-pixel
additions reproduce the earlier definition within 7.11e-15 electrons/second.

## Repeat the learned operations

Three arms use the same nominal data and masks:

- Local primary: replay the old own-cutout injection.
- Detector-consistent: inject all 172 sources wherever they illuminate the
  retained cutouts.
- Donor only: inject only Gaia 1622915828734016128, the previously identified
  reference-neighbour donor, while every other source has zero input signal.

All 172 science profiles are relearned, with independent per-exposure source
flux and response-aware sky fits. Both aperture increments are recomputed;
science aperture-ratio transport and reference precision are refitted. The
original source-fold/two-day exclusions apply to ratio fits. Reference
precision still uses full histories, as in the current candidate.

The five-source pilot includes both cross-illuminated science targets, the
reference donor and two other controls. It replays bitwise in the full run.
Nominal fitted fluxes reproduce exactly; the earlier local-injection fluxes
reproduce to maximum 1.11e-15 fractional difference. The old final paired
light curves and injected ratio prediction also replay within declared
numeric tolerances. Global WCS, response, profile predictors, photon means and
physical count-loss predictions remain fixed. No new photon/readout noise is
drawn; this closes spatial injection coupling, not the entire detector chain.

| Test | Local primary | Detector-consistent |
|---|---:|---:|
| Worst amplitude error across all 172, % | 0.48120 | **1.55152** |
| Number failing the fixed 0.5% criterion | 0 | **1** |
| Recovery-error RMS median/p95, ppm | 1.692/3.375 | 1.729/3.559 |

The failing source is Gaia 1626038957153297152 (science index 71), beside
Gaia 1626038201239063936 (index 70), separated by 10.88 pixels. Their signed
amplitude errors change 0.10390%→1.55152% and 0.05617%→0.35557%, respectively.
The modest median change hides a real individual failure.

![All-source amplitude gates and donor-only imported signals](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/source_consistent_refit_full/calibration/signal_gates.png)
![Actual known and recovered signal histories](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/source_consistent_refit_full/calibration/recovered_histories.png)

The donor-only 1-ppt control has a maximum false-signal RMS of 0.569 ppm among
the other 171 science stars (95th percentile 0.295 ppm). Most stars have
numerically zero response because their reference weights exclude the affected
reference. The donor's own amplitude error is 0.000040%. These are controlled
responses to this injection, not measured real-neighbour contamination or a
general precision floor.

## Localize the failure before changing the correction

The learned normalized stamp is

$$Y_{t,p}=\frac{D_{t,p}-B_{t,p}}{A_t},\qquad
P_{t,p}=\sum_v H_{t,v}Y_{v,p}.$$

Here $D$ is the actual locally injected calibrated pixel rate, $B$ the fitted
response-aware sky rate, and $A$ the measured radius-5 aperture rate, all in
electrons/second. $Y$ and the learned profile $P$ are dimensionless. $H$ is
the existing dimensionless exposure-mixing operator with the target exposure
excluded, and $v$ indexes training exposures. Its inherited explanatory fields
are fixed in this test.

The profile response separates as

$$\delta Y_{t,p}=\frac{\delta D_{t,p}}{A_t}
-\frac{\delta B_{t,p}}{A_t}
-Y_{t,p}\frac{\delta A_t}{A_t}.$$

The deltas denote the known additional neighbour pixel signal and its induced
sky/aperture changes. Each term is dimensionless. Propagating these through
$H$ and the weighted flux fit separates direct pixel fitting, profile pixel
data, profile sky subtraction and profile aperture normalization. Spatially,
the same perturbation is separated into original aperture pixels, sky-ring
pixels and the remaining retained gap pixels.

![Underlying injected neighbour pixels and fixed region masks](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/source_consistent_refit_full/attribution/injected_pixels.png)
![Pixel-region and estimator-pathway decomposition](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/source_consistent_refit_full/attribution/pathways.png)

For source 71, the full nonlinear coupled-minus-local flux change has
14.8416-ppm RMS before reference calibration. Sky-ring perturbations produce
14.8763 ppm, while aperture-region perturbations contribute only 0.03472 ppm
and partly cancel. The profile-aperture-normalization pathway contributes
14.8428 ppm; direct fitting and the profile pixel-data pathway nearly cancel.
The independent source-70 control shows the same structure at 3.2118-ppm RMS.
These correlated contributions are not independent variances and must not be
added in quadrature.

The derivative decomposition agrees with the actual full nonlinear refits to
maximum 3.91e-5 ppm for source 71 and 1.21e-6 ppm for source 70, far below the
predeclared 0.05-ppm linearization tolerance. This establishes the mechanism
for these known perturbations. It does not establish the true amplitude or
shape of the real neighbours.

## Required residual evidence and provenance

All absolute and cell-relative colour × magnitude × 3×3 CCD grids were
generated and inspected for all three pilot and full arms. They use the same
original masks, old-r5 normalization and four snapshot FFIs 1/601/698/1225.
The pilot has four eligible input stars (three in canonical plotting range),
so no bin reaches MIN_N=5 and the panels are honestly blank. The full grids
have 165 eligible inputs, 159 in canonical range; sparse bins remain blank.
Canonical colour scales are automatic per image. The common-scale injected
pixel panels above are the appropriate amplitude comparison for the failure.

The [existing shared gallery](http://localhost:18888/residual-grids/resgrid.html)
preserves all 127 prior entries and adds six labels (133 total):
`raw_signal_{pilot,full}_{local_primary,detector_consistent,donor_only}_20261004`.
Its publication manifest and image hashes are in
`results/source_consistent_residual_grids/publication.json`. The external
Claude Artifact mirror was unavailable and was not republished.

Frozen commands and exit records are in `condor/source_consistent_prepare`,
`source_consistent_refit_pilot`, `source_consistent_refit_full`,
`source_consistent_calibration`, `source_consistent_grids`, and
`source_consistent_attribution`. Full refit Condor job 2429.0 on schedd
plscience3/execute plscience2 exited 0; all owned jobs are terminal. Input
hashes are recorded and verified before refitting. No production inputs or
nominal candidate light curves changed.

**Conclusion / next:** the stronger test overturns a blanket interpretation
of the earlier amplitude pass. The demonstrated defect is neighbour light in
the sky-based aperture normalization. Next protect that normalization using
catalogue-defined neighbour-aware sky support or an independently constrained
neighbour/sky model, preserving primary aperture support and all original
science IDs. Compare actual photometry, conditional noise, shared trends and
these same source-consistent injections. Do not subtract the known injection,
fit away a science trend, relax the gate or drop the failing star. Detector
calibration, WCS, broader populations and photon-noise validation remain open.