> **2026-10-04 follow-up:** [SOURCE_CONSISTENT_SIGNALS.md](SOURCE_CONSISTENT_SIGNALS.md) completes spatially consistent injections/refits. The stronger test fails at1.55152% amplitude error; controlled attribution identifies sky-ring-driven profile normalization. Earlier local-only numerical results remain historical/conditional.

> **2026-10-03 follow-up:** [NEIGHBOUR_LEAKAGE.md](NEIGHBOUR_LEAKAGE.md) finds actual source/reference pixel overlap and documents that earlier primary injections were local to each science cutout. Their amplitude results remain conditional; no complete source-consistent signal pass is claimed.

# Lightly masked reference expansion — 2026-10-03

**Conclusion: the tested expansion is not adopted.** Adding 47 references
to the original 235 does not remove the two persistent group trends. It
slightly increases imported reference variability and fails the existing
0.5% primary-amplitude criterion for one of 172 science sources. The
instrument-free/photon-limited goal remains unmet. Calibration accuracy,
neighbour leakage and physical noise accounting remain open.

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/light_*.py`.
The [existing results canvas](http://localhost:18888/temporal-raw-epsf/canvas/temporal_raw_epsf.html)
contains this milestone with the preceding evidence. Code is not pushed.

## Selection without science-flatness tuning

The previous static-shape and population-prior experiments failed, so this
pilot retains the original population PSF and measured temporal WCS/shape.
Among 2,522 candidates, require complete original radius-2 support, nominal
radius-3 aperture inside the CCD, finite positive modeled response, and at
most 0.005 fractional response departure between retained and complete
radius-3 support. This threshold was fixed before viewing science curves.
The same observed outer ring and opposite-orbit response-aware sky weights
are used for both model projections. No missing pixel is filled with data.

![Selection and CCD coverage](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_inventory/selection.png)
![Actual retained masks and model responses, not observed hidden pixels](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_inventory/selected_masks.png)

All 51 selected new candidates pass on all 1,224 original quality-zero
exposures. A previously used broad 20-ppt reference-only variability screen
removes four, leaving 47 provisional references: 9/7/13/18 in readouts
A/B/C/D. Their variability is unresolved, not certified astrophysical.
All 51 histories remain saved. Heavily masked, fainter and crowded stars
remain in the larger objective.

## Match the physical correction to the actual estimator

The extraction uses

$$r_{s,t}=\frac{\sum_p w^{\rm partial}_{s,t,p}P_{s,t,p}}
{\sum_p w^{\rm complete}_{s,t,p}P_{s,t,p}},\qquad
\widehat F_{s,t}=\frac{\sum_p w^{\rm partial}_{s,t,p}
(I_{s,t,p}-\Delta I_{s,t,p})}{r_{s,t}}.$$

Here $s$ indexes a reference source, $t$ an exposure, and $p$ a stamp pixel;
$P$ is the inherited dimensionless population PSF at the measured position
and temporal shape; $w^{\rm partial}$ and $w^{\rm complete}$ are dimensionless
radius-3 aperture/sky weights on retained and hypothetical complete support;
$I$ is the measured calibrated pixel rate in electrons/second; $\Delta I$
is the simulated signed onboard-count increment in electrons/second; $r$
is the dimensionless model response ratio; and $\widehat F$ is the resulting
model-normalized aperture rate in electrons/second. The complete weights
are used to define the model response, not to claim missing pixels were observed.

The earlier six-frame pilot projected physical increments through complete
weights. That is retained only as a sensitivity: the actual correction must
use the partial weights and divide by the same response as the data. Their
full-cadence difference has median/p95 temporal RMS 5.91/22.69 ppm for the
47 new sources. This is a modeled operator difference, not an error bound.

The full physical job `2414.0` on schedd `plscience3.stsci.edu` completed
normally with exit 0 on `plscience2.stsci.edu`: 16 CPUs requested, 854 MB
recorded memory, 690 seconds execution. It used the same eight-replicate
isolated-source recipe and fixed means/seeds as the pilot. All 47 sources
have all 1,224 exposures. Every shared pixel prediction on the six pilot
exposures is bitwise equal. Old restored references reproduce the current
235-source input within 2.22e-16 relative difference. Old science response
features replay exactly; original reference features differ by at most
3.33e-16.

![Observed candidate histories and matched physical correction](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_restore/reference_histories.png)
![Underlying simulated pixel changes, common scale](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_restore/physical_pixel_changes.png)

New-reference count-loss temporal RMS is median/p95 0.2464/0.4251 ppt.
These are simulated contributions under approximate photon means and
profiles. The historical scalar simulation standard error uses a different
aperture operator; it is not assigned to the present projection.

The first restoration gate incorrectly compared aggregate fingerprints
that include different selected source/epoch sets. It failed before writing
results. The corrected gate verifies shared predictions, seeds, fluxes,
geometry, physical settings and frozen renderer hashes directly. Both
attempts remain recorded; no physical simulation was rerun to force agreement.

## Fixed-cohort comparison and paired signals

Keep all 172 original science stars, their pixel fits, 165 eligible stars,
354 magnitude/colour/spatial/readout groups, and original source-fold/two-day
ratio fits. Evaluate those same fitted coefficients at each added reference's
own location/features. Compare original 235 versus expanded 282 with the
same fixed k=80 rule. Reference precision is re-estimated from full histories;
it is not a time-held-out precision test. No science trend chooses a weight,
threshold, new coefficient or winning method.

| Diagnostic | Original 235 | Expanded 282 |
|---|---:|---:|
| Ten-star spatial group six-hour RMS, ppt | 0.12332 | 0.12858 |
| Sixteen-star neighbourhood six-hour RMS, ppt | 0.18360 | 0.17859 |
| Worst primary amplitude error, % | 0.48120 | **0.53318, fails** |
| Primary recovery RMS median/p95, ppm | 1.692/3.375 | 1.760/3.467 |
| Imported reference variability RMS median/p95, ppm | 86.67/119.05 | 86.90/120.77 |
| Matched reference-half difference RMS median/p95, ppm | 102.22/121.25 | 113.61/138.35 |

![Group histories and reference sensitivity](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_compare/comparison.png)
![Actual underlying science histories, fixed cohort and shared scales](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_compare/individual_histories.png)
![Actual paired primary signals and amplitude gate](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/light_reference_compare/primary_signal.png)

The failing source is Gaia 1643404334564563712: signed amplitude error
changes from −0.48120% to −0.53318%. The injected profile is the previously
recorded independent analytic profile; the matching refitted science-profile
and ratio results are reused exactly. WCS/global response/photon means remain
fixed. Reference-variable injections have 1-ppt amplitudes and 5/10/20/30-day
periods. Original phases are unchanged when new stars are appended. The
reference halves use catalogue-order parity, preserving the original prefix;
this differs from other historical partitions, so compare only this matched
pair. No instrumental significance is assigned to these descriptive changes.

Exact commands, arguments, snapshots and exit records are in `condor/` jobs
`light_reference_inventory`, `light_reference_full`, `light_reference_features`,
`crm_light_reference_pilot`, `light_physical_gate`, `crm_light_reference_full`,
`light_reference_restore`, `light_reference_restore_checked`,
`light_reference_compare`, `light_reference_signal_evidence`, and
`light_reference_signal_evidence_colours`. Results share the script names
under `results/`, except the restoration's successful result is
`results/light_reference_restore/`. The empty orbit bins remain NaN and
produce expected all-NaN plotting/reduction warnings; no epoch is imputed.
No new PSF training or fixed-scene fit was performed in this milestone, so
the shared training-residual gallery is unchanged.

**Conclusion / next:** this particular reference expansion is rejected as a
replacement; it does not clear upstream calibration or missing-profile
errors. Before another expansion or correction, quantify real-neighbour
leakage using catalogue geometry, independent profile families and variable
neighbour injections. Then test corrections using retained pixels and held-out
sources/epochs. The detector-gain, chromatic response, background, WCS and
physical covariance questions remain open; the current candidate is unchanged.