# Disjoint-reference engineering response — 2026-10-03

**Conclusion:** learning the response from adjacent-exposure changes nearly
eliminates the tested leakage of reference-star variability, but leaves more
shared slow structure in real light curves. The present linear engineering
model cannot replace full reference histories. No correction is promoted;
instrument-free, photon-limited photometry remains unachieved.

Run: `/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/`.
Code: local branch `codex/temporal-raw-epsf`, worktree
`/home/kshukawa/syndiff_pipeline/dev/.worktrees/temporal-raw-epsf`.
Scripts: `temporal_raw_epsf/reference_engineering_response.py`,
`reference_engineering_assessment.py`, and the standalone executable
`test_reference_engineering_response.py`. Code remains local.

## Question and method

The current ensemble correction imports actual reference histories at unit
fractional amplitude. Those histories carry both instrumental response and
their own astrophysics. This test replaces the full histories with a response
to measured jitter, camera temperature, local sky and WCS motion. It differs
from the previous global-fast experiment, which trained on science amplitudes.
No science amplitude trains the present throughput response.

For a reference star, use

$$y_{r,t}=\ln\!\left(A_{r,t}/A_r^0\right)+R_{r,t},
\qquad y_{r,t}=\mu_r+\mathbf{x}_{r,t}^{\mathsf T}\boldsymbol\beta+\epsilon_{r,t}.$$

Here $r$ labels a disjoint reference and $t$ an exposure; $A$ is its
physically count-corrected r3 rate in electrons/second; $A^0$ is its fixed
sector median in the same units; $R$ is the dimensionless predicted relative
log r5/r3 aperture response at that reference's actual properties and position;
$y$ is dimensionless relative log flux; $\mu_r$ is a constant nuisance offset;
$\mathbf{x}$ contains the 31 engineering/catalogue interactions, made
dimensionless using training RMS scales inside the regression;
$\boldsymbol{\beta}$ is the shared fitted dimensionless response in that scaled convention;
and $\epsilon$ contains stellar variability, noise and missing instrumental
effects. Saved coefficients are converted back to the original feature units.

The three predeclared fits use levels, adjacent differences and 12-cadence
(six-hour) differences. The fit removes each training star's constant or
mean difference. Rapid pairs are disjoint, do not cross cadence gaps, orbit
boundaries or two-day blocks, and exclude the evaluated block plus a 1.5-day
guard. Constant stellar slopes therefore do not identify a rapid response.
All arms use the same pooled Huber weighting, six iterations and existing
scaled ridge of 0.001. No nonlinear response or tuning to science flatness
is introduced. Collinearity and long-term extrapolation remain limitations.

Crucially, aperture-ratio transport is refitted inside the same outer
science-fold/time exclusion. Previously cross-fitted reference ratios are
not reused in a way that could bring the held block into training. Reference
halves are assigned with seed 821937, balanced within each readout. Their
sizes are 119 and 116. Each predicts the other; their predictions for a
science source are averaged. Populations stay at 235 references, 172 science
sources and 165 eligible science sources, across 1,224 FFIs.

A matched full-history control uses the same outer fits, guards, halves and
injection bookkeeping. It averages 40 local references from each half and
estimates their precision from training epochs only. Reference fluxes at the
evaluation exposure remain legitimate inputs for this history-based control.
A static r3 median and zero-mean aperture-ratio field give a consistent
normalization across outer blocks. The previous k80 correction is retained
as an additional benchmark; it is not identical to this nested control.

## Validation and measured results

Known-response tests recover the synthetic coefficients within 0.2%; changing
excluded source/time values leaves coefficients exactly unchanged. Tests also
verify rejection of cadence/orbit gaps and constant slopes in rapid fits.
Primary pixel injections recompute affected flux normalization, aperture
transport and response coefficients. Inherited pixel WCS/shape and physical
count predictions remain fixed, so this is not a full physical-model injection.
Separate 1-ppt sinusoidal reference injections use 5–30 and 45–120-day periods
and fixed phases; their import is measured after refitting the correction.

| Method | Six-hour ensemble RMS [ppm] | Median import, 5–30 d / 45–120 d [ppm] | Median half-disagreement [ppm] | Worst primary amplitude error |
|---|---:|---:|---:|---:|
| Nested reference histories | 81.18 | 89.13 / 44.55 | 165.38 | 0.332% |
| Engineering levels | 137.67 | 79.83 / 66.81 | 211.01 | 0.581% |
| Adjacent engineering differences | 135.35 | 0.762 / 0.051 | 78.65 | 0.581% |
| Six-hour engineering differences | 177.43 | 14.76 / 0.697 | 136.62 | 0.582% |

The ensemble RMS includes astrophysics and calibration covariance, not just
instrumental errors. Injection and half-disagreement statistics cover all
172 science targets; ensemble values use the fixed 165 eligible sources.
The chosen reference signals are sensitivity tests, not a confidence interval
on the real reference population. A smaller half-disagreement alone cannot
prove that the common response is correct.

![Primary signal recovery and actual worst-source error history](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/phot_full_pixel_local_crm/reference_engineering_assessment/primary_injection_validation.png)

The worst adjacent-arm source's recovered-minus-injected history has a
localized error near the late-first-orbit jitter episode, visible even though
the full recovered sinusoid closely overlaps the injected one. The amplitude
summary alone does not describe every temporal error. This figure was added
by job 1137 (exit 0, 12 seconds) without changing fits or numerical outputs.

![Reference contamination and independent-half disagreement](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/phot_full_pixel_local_crm/reference_engineering_assessment/reference_contamination.png)

![Actual reference histories underlying the held-out check](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/phot_full_pixel_local_crm/reference_engineering_assessment/heldout_reference_histories.png)

The held-out adjacent response leaves a slow decline visible in both
reference halves. Display limits of ±3 ppt do not exclude any sources from
the numerical outputs. This is evidence that the chosen response basis
does not capture everything in those histories, not an identification of
which remaining component is instrumental.

## Fixed multiscale audit

All 354 preexisting groups are tested: magnitude, Gaia BP−RP colour, multiple
spatial scales, neighbours and readout regions. The same 4,096 whole-history
sign randomizations and seed 710023 reproduce the old baseline's adjusted
p values exactly. This conditional null assumes independent sign-symmetric
stellar histories. Shared calibration covariance, asymmetric variability and
mixed-sign physical responses remain limitations. The invalid slow-signal
phase null is not used as a new acceptance criterion.

| Method | Sign-family flags / 354 | Preselected 16-star neighbourhood RMS [ppt] |
|---|---:|---:|
| Existing k80 reference correction | 1 | 0.1948 |
| Matched nested histories | 1 | 0.1852 |
| Engineering levels | 6 | 0.2728 |
| Adjacent engineering differences | 10 | 0.2702 |
| Six-hour engineering differences | 23 | 0.2925 |

The original neighbourhood remains flagged in the matched history control,
with adjusted p=0.0334. Counts are associations under this null, not counts
of instrumental causes. Adjustment is within each predeclared arm, not across
all model choices. No flag disappears because a source or group is dropped.

![Same group histories under each method](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/phot_full_pixel_local_crm/reference_engineering_assessment/group_comparison.png)

![Measured individual science histories behind the engineering comparison](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/phot_full_pixel_local_crm/reference_engineering_response/individual_group_histories.png)

![Measured individual science histories for the matched history control](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/phot_full_pixel_local_crm/reference_nested_history_control/individual_group_histories.png)

All figures were inspected. The stronger residual structure rules out these
three linear engineering replacements as a complete solution. It does not
rule out reference-star contamination as one contributor, unknown detector
calibration, nonlinear/hysteretic optical response, or better independent
reference populations. Full histories retain useful time-dependent information
that this predictor set misses, along with unwanted stellar variability.

## Provenance and next discrimination

Frozen jobs on plscience3: engineering 1100 (8 CPUs/8,000 MB, exit 0, 67 s),
matched histories 1110 (same request, exit 0, 50 s), assessment 1119
(4 CPUs/4,000 MB, exit 0, 30 s). Source snapshots and exact commands are under
the run's `condor/` directories. Data and numerical predictions remain on
`/astro`, and the existing candidate is unchanged.

Next investigate missing physical common responses and independent reference
coverage rather than adding flexible per-star trends. One additional physical
check is photometric observer-motion response using the WCS agent's actual
TESS ephemerides; distinguish an observer boost from source orbital beaming
before applying any coefficient. This is only a proposed diagnostic, not a
detected cause or a correction. The uncorrected mean-image/scene and hp_d
offset work, broader stellar coverage and full noise budget also remain open.