1. What has been removed, and what is left?

The original centroid fits did not estimate an additional local background; they relied on the prepared images. This makes a residual-background test necessary.
We inspect calibrated FFIs, saved large-scale background (ks_b), saved Hotpants background (hp_b), and centroid input (hp_d) separately. Their differences are not assumed to be pure sky or an exact reconstruction of the historical recipe.
The “first” samples occur 2.5 hours after science observations begin; “last” samples occur 2 hours before they end. One interior reference is used separately for each orbit.

Actual source-masked 32×32 pixel medians across the detector. Colour scales are fixed within each column across all six epochs; white blocks have no accepted pixels. Broad startup illumination appears in the FFI and background estimate; the centroid-input map uses a much narrower scale.

Robust quadratic surfaces fitted to masked 5–15 pixel annuli around 168 selected sources; 150–151 usable per epoch. The inset exposes small residual central levels. These surfaces mix possible sky errors with wings/neighbour/template residuals.

Actual centroid-input stamps and accepted annular pixels. Star and annulus panels use different explicitly labelled scales, each shared across epochs. Gray means excluded, not zero. Source-shaped structure remains within nominal background support.

The same masked detector pixels at each endpoint minus its own orbit-interior image. No recentering or star-model subtraction: source motion, PSF changes and background can all contribute.

Finding: no strong endpoint-only elevation appears in median local gradient magnitude on this cohort. That does not establish a pure-sky measurement or zero centroid bias.

2. Refit the same stars with different local backgrounds

Frozen saved ePSFs; identical 11×11 native pixels and weights for zero, constant and planar background arms. Positions and amplitudes are free; the ePSF itself is unchanged.
168 selected isolated stars with Tmag 9–12.8; 133 retain adequate valid support at every epoch. The 35 excluded sources are not validated. Raw-image convergence does not select the primary subtracted-image cohort.
The metric is the change in the plane-minus-zero correction from each orbit's interior exposure. Static biases and inter-orbit jumps are removed by this definition.

Paired temporal background-model sensitivity. Top: actual centroid-input images. Bottom: raw calibrated FFI control, for which a zero-background model is deliberately inadequate; raw differences also include crowding/template effects. Signed medians are distinct from individual-star scatter.

Actual per-star plane-minus-zero correction changes, enlarged with a common 0.5 millipixel vector key. These are centroid shifts in detector coordinates, not the opposite-sign sky-map displacements used in the orbital atlas.

Underlying data, fitted nuisance plane, and data-minus-PSF residuals for strongest and representative sensitivities. Background and residual panels have their own stated scales; small smooth planes do not remove all profile residuals.

PDF · Full-resolution PNG

Window	n	Median x [mpx]	Median y [mpx]	Star RMS [mpx]
Orbit47 start sample	133	0.008	-0.004	0.337
Orbit47 end sample	133	-0.006	-0.054	0.304
Orbit48 start sample	133	0.020	0.032	0.349
Orbit48 end sample	133	-0.004	0.011	0.236

Interpretation limit: the new optimizer’s zero-background positions differ from the saved centroid solution (about 1.5–2.1 millipixel median on full support). An original-path 10-star audit reproduces the archive; our paired sensitivity remains an alternative-fitter result, not a literal correction to the archived WCS. A frozen ePSF may already contain upstream background bias.

Original-path reproduction and alternative-solver audit

Ten stars in the first exposure: the original photutils procedure reproduces archived coordinates, while the alternative optimizer gives slightly different positions with either extraction centre. This is numerical consistency, not an astrometric accuracy test.

Actual images, original-path fitted profiles and residuals behind the ten-star reproduction control; stellar residuals share a 3% peak scale.

3. Add known backgrounds while keeping the true position fixed

Two preselected detector locations and their six saved profiles; brightness scaled to Tmag 9, 11, 13. No noise, neighbours or ePSF retraining: these are conditional shared-profile controls.
Source-matched annular quadratic fits set the injected backgrounds. “Measured plane” retains only that quadratic fit’s constant and linear terms; it is not an independently known sky surface.
At Tmag 13, the full quadratic injections give median 0.074 and maximum 0.255 millipixel endpoint-relative bias without a fitted background, across eight endpoint/source pairs. A fitted plane leaves median 0.009 and maximum 0.036 millipixel here.

Known-position bias changes from each orbit interior, separated from static centroid offsets. Shading spans the selected profiles/endpoint cases; it is not a population confidence interval.

Underlying frozen stellar profiles, added backgrounds, injected images, and residuals under zero/constant/plane fits. Background and residuals share their scale within each row; different rows may use different scales.

Brightness and gradient-direction controls

The same additive background becomes more important for fainter sources. A constant term cannot absorb a gradient; a plane cannot fully absorb curvature. Ideal recovery here is conditional on the identical generating and fitted profile.

PDF · Full-resolution PNG

Induced signed shifts for rotated gradients on the two selected profiles. Direction and profile shape matter; these two locations do not bound every star or detector region.

PDF · Full-resolution PNG

4. What this establishes—and what it does not

Could background removal bias orbital WCS measurements?

Yes in principle; the direct additive effect is small in this bounded C 3/CCD 3 test. This does not certify every exposure or rule out errors inherited by the trained ePSF.
Scope: Sector 20 camera 3 CCD 3, six exposures:2.5 hours after each orbit's science start, one interior exposure per orbit, and 2 hours before each ending. These are representative snapshots, not complete first/last-day coverage.
Our subtraction cannot alter the independently produced mission-header WCS. It can affect our image centroids and trained WCS. Mission processing may have its own background sensitivity, not tested here.

What was actually fitted

Historical centroid ECSV metadata records no additional local background estimator. The temporal fitter likewise used prepared images without a nuisance background column. The modern single-image fit uses a different image generation; its upstream harmonic background processing must not be confused with the historical lane.
The old lane has saved background products but no frozen subtraction recipe establishing every historical parameter. Raw-minus-hp_d includes removed sources/templates; it is not a pure background estimate.
New positional test:168 spatially selected isolated sources, Tmag 9–12.8;133 have sufficient common native-pixel support. All attempted fits converge. The 35 excluded sources are not certified.
The same saved ePSF,11×11 pixels, uniform weights and fixed valid-pixel mask are used for zero, constant and sloping-plane background fits. Source positions and amplitudes are refitted; ePSFs are not retrained.
Report the change in each star's background correction relative to its own orbit-interior exposure. This removes static offsets and does not test the inter-orbit jump.

What the images show

The large background patterns, including extra broad light at the second startup snapshot, appear in the calibrated FFI and saved background estimate. A comparable broad enhancement is not apparent in the subtracted centroid-input map.
In the 168-star annulus diagnostic,150–151 sources are usable per exposure and 150 are common to all six. Median residual gradient magnitudes range 0.0246–0.0288 electron/s/pixel, without a strong endpoint-only elevation.
Masked annuli still contain arcs and compact residuals. Sky background, stellar wings, neighbours and template errors remain mixed; a fitted surface is not independently known sky.

What changes the positions

Allowing a plane changes endpoint-minus-interior corrections by 0.236–0.349 millipixel RMS across the 133 stars. The corresponding median vector is below 0.055 millipixel in each endpoint snapshot. A millipixel is 0.001 native detector pixel.
These quantities describe sensitivity within the new fitter, not absolute positional accuracy. No projection onto the thermal WCS modes was fitted, and the star and WCS-probe populations have different denominators; no explained fraction is claimed. Spatial maps and outliers accompany the medians.
The new optimizer's zero-background solution differs from archived centroids at roughly 1.5–2.1 millipixel median even on full 121 pixel support. A separate 10-star audit using the original photutils path reproduces the saved positions to numerical precision; the new-solver difference is retained as a limitation, not attributed to bad saved products.
Changing a local background term cannot undo errors already absorbed into ePSF shape/origin, source removal or template construction. That end-to-end pathway remains untested.

Controlled injections

Two preselected detector locations, their six saved profiles, Tmag 9/11/13, known fixed positions and source-matched annular background surfaces;2052 fits, no noise or neighbours and no ePSF retraining.
AtTmag 13, full measured quadratic surfaces produce a median endpoint-relative bias 0.074 millipixel and maximum 0.255 millipixel when no background is fitted, across eight endpoint/source pairs. Fitting a plane leaves median 0.009 and maximum 0.036 millipixel in this shared-profile control.
Stronger injected gradients can produce millipixel shifts. The control demonstrates the mechanism and scales for these profiles; it is not proof that all real backgrounds are this small.

Paper consequence and next decisive test

The sampled direct residual constant/gradient is unlikely to be the main source of the several-millipixel orbital remainder. This is a constrained negative result for one CCD and these sources/exposures.
Keep background-induced profile/centroid bias in the physical-cause budget. Thermal correlation alone cannot rule it out.
To close the remaining pathway, rebuild backgrounds and ePSFs from the same raw endpoint images under alternative source masks/background models, then score fixed held-source positions and residual images. Expand to other detectors and the earliest science exposures.

Conclusion: broad background removal appears effective in the inspected images, and direct local-background sensitivity is small for the tested cohort. Background effects inherited by the ePSF or upstream subtraction remain a live uncertainty; no blanket all-camera or full-day certification is justified.

Download findings and validation bullets · Figure hashes