# Native RAW phase support: independent stars versus repeated exposures **2026-10-05 conclusion:** the fixed independent-star pools cover nearly all quarter-pixel phase bins, whereas an individual source's modelled coadd centre moves over only a small fraction of a pixel during this sector. Phase coverage does not eliminate the known spatial/colour mismatch, especially for the two red targets. These results guide response reconstruction; they do not prove identifiability or independent WCS accuracy. Real photometry is unchanged and the full instrumental-systematics/photon-noise goal remains unmet. Run: `/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001`. Code: `/home/kshukawa/syndiff_pipeline/dev/.worktrees/temporal-raw-epsf`, branch `codex/temporal-raw-epsf`. Scripts: `pooled_phase_inventory.py` and `temporal_phase_support.py` under `temporal_raw_epsf/`. No push or merge. ## Fixed independent-star pools Read the existing 64-source pools from `results/pooled_positive_profile_fine/`. Membership and source/overlap exclusions are unchanged. Recompute the exact per-frame positions used by that fit: saved reference positions plus archived WCS temporal increments and opposite-source-fold RAW residual fields. Pixel phase is the fractional part of each stamp centre; integer stamp origins do not change it. Divide each axis into four equal phase intervals. Every target/epoch pool occupies 15 or 16 of the 16 bins. Some occupied cells have only one contributing source. The counts and kernel mass differ, so 64 training stars do not imply 64 equally weighted constraints or four-times effective resolution. The stars are different physical sources, and their PSFs need not be identical across position, colour or brightness. ![Actual phase-bin counts, kernel mass and target phase](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_coverage.png) The phase-resolved colour mismatch is substantial for red targets 138 and 358, consistent with the preceding convex-hull support audit. It is not fixed merely by populating more subpixel phases. No source selection or extrapolation correction is introduced here. ![Training-minus-target colour by phase](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_colour.png) For the underlying images, subtract the inherited scalar sky from each RAW stamp and divide by that exposure's measured fixed-r3 aperture rate. The display is the median of available native pixels in each phase bin; missing pixels remain blank. There is no subpixel resampling, recentring or temporal detrending. This display normalization differs from the fitter's historical reference normalization with free source amplitudes. It is not a measured instantaneous optical PSF. Phase-bin membership can change between epochs, so differences between these displays do not isolate temporal PSF changes. ![Measured training stamps for target 134](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_stamps_134.png) ![Measured training stamps for target 138](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_stamps_138.png) ![Measured training stamps for target 359](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_stamps_359.png) ![Measured training stamps for target 358](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_stamps_358.png) ![Measured training stamps for target 191](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_stamps_191.png) ![Measured training stamps for target 190](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/pooled_phase_inventory/phase_stamps_190.png) ## Repeated exposures of individual sources Reconstruct the same modelled coadd-centre positions across all 1,224 original FFIs for all 431 original extraction positions. The two-epoch positions replay the phase inventory within an absolute tolerance of 1e-13 pixel. This is a numerical convention check, not new centroid validation. Keep the complete time series and population, including orbit edges. | Auxiliary index | x span [pixel] | y span [pixel] | Occupied original 4×4 phase bins | |---|---:|---:|---:| | 134 | 0.03787 | 0.06300 | 2 | | 138 | 0.03785 | 0.06301 | 2 | | 359 | 0.03884 | 0.05794 | 1 | | 358 | 0.03744 | 0.05690 | 2 | | 191 | 0.03853 | 0.06229 | 1 | | 190 | 0.03755 | 0.06112 | 2 | These are peak-to-peak spans of the inherited WCS-predicted exposure-average positions, not instantaneous pointing excursions, measured astrometric errors or new velocity-aberration/thermal coefficients. For display the phase origin is rotated to put the first exposure at (0.5,0.5), avoiding spurious long tracks at a modulo-one wrap. The reported bin occupancy uses the original unrotated detector phases, so it is distinct from the illustrative grid. ![Individual modelled centre tracks through every exposure](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/temporal_phase_support/target_tracks.png) Across all 431 positions, median/p95 x spans are 0.03628/0.04229 pixel and y spans 0.05872/0.06327 pixel. Individual sources occupy one to four original quarter-pixel phase bins. These small mean-position spans do not rule out additional information from changing within-exposure jitter kernels, physical model constraints or shared stars. A rank/information calculation is still needed for any proposed reconstruction; bin occupancy alone is insufficient. ![Complete population coverage by position and magnitude](/astro/armin/koji/syndiff/dev_runs/temporal_raw_epsf_20261001/results/temporal_phase_support/population_support.png) ## Execution and next step Both frozen local jobs, `condor/pooled_phase_inventory/` and `condor/temporal_phase_support/`, exit 0 on plscience3. All ten figures were inspected. Source IDs, pool weights, per-pixel counts, stacks and modelled tracks are saved in their corresponding `results/` directories. The local two-CPU / 2,000-MB resource hints are recorded but memory limits are not enforced. No new ePSF fit occurs in these two inventory jobs. The separate directional-wing fit uses the same six targets and 64-source pools at FFIs 1 and 601. Its results and convergence must be checked before judging whether the enlarged positive basis helps held target pixels or same-exposure flux consistency. Keep colour/spatial extrapolation and the difference between coadded and instantaneous profiles explicit throughout. **Conclusion:** multiple independent stars supply phase diversity unavailable from each star's mean-position track alone, but transferring profiles across stellar properties remains a material limitation. This inventory does not establish a recovered subpixel response or corrected light curves.