@@ -26,6 +26,7 @@ Next: pull, bundle with output_sed, compare magnitudes (µJy fluxes) against the
 
 ## Log
 
+- 2026-09-17 — result — Witt–Wynne SIEP projection now a catalogue product (pipeline PR #86 merged 2026-09-17, stage 7 of the inspection bundle). Ran catalogue/scripts/witt_wynne.py on the ten pulled dr1_prelim_grade_ab tiles: 10/10 rows, 0 skips, source centre recomputed from the fit (pre-source-key wcs.json). Verdicts: 9 tiles 2-image, 1 tile 4-image; PointSolver on the full SIE+shear models agrees on 9/10 (Tile102007903 says 2 where the full model gives 4). Positions differ from the full model by 10–40 % of b; caustic-matched projection is the default because the vector sum disagrees in PA by 15–84° on 6/10 tiles. Redshifts are placeholders (z_l 0.5, z_s 1.0) so lags are fiducial. Files: inspect/dr1_prelim_grade_ab/witt_wynne.csv + per-tile witt_wynne.in. Recorded by the agent at /prm close-out of #84; human to confirm wording.
 - 2026-09-12 — run — 342796_[1,3,8] submitted: ordered_rerun_catalogue_parity: CPU SED chain for the three tiles still missing one, via a direct ssh sbatch --array=1,3,8 on hpc/batch_cpu/submit_sersic_waveband (md5 65db7d8b… on RAL = the clone's commit 88dda9d). Tiles 1 (Tile102007299…) and 3 (Tile102007903…) had no SED chain at all because 342629_1/_3 only finished vis_lp/vis_pix on 2026-09-12 06:58 and 08:50 (walls 1-09:11:56 and 1-11:04:44); their initial_lens_model/vis_lp zips were seeded into output_sed by cp -p (md5-identical, sources intact) so stage 1 short-circuits. Tile 8 (Tile102008532…) is a resume: 342648_8 TIMEOUT at 8:00:19 on the GPU with vis, nir_y, nir_j and nir_h written but decam_g hung after its jit compile completed at 15:58:39 and produced zero Nautilus iterations in the remaining 7h45m (no checkpoint.hdf5; only pre-fit files in the search dir), leaving decam_r/i/z unrun. Nothing deleted on RAL — the completed band zips short-circuit and the empty decam_g dir restarts as a new search. Tile 0 (Tile102005065…) was NOT resubmitted: its 5 zips are a complete chain, because band membership is discovered from the cut-out FITS and that tile carries no DECam imaging (342648_0 exited 0 straight after nir_h).
 - 2026-09-11 — note — **CPU verification PASSED and was faster than the GPU.** `342668_9` (ral, 8 CPU, JAX on the CPU backend) COMPLETED in **9:24** wall, exit 0, all eight `sersic_lens_model/{vis,nir_y,nir_j,nir_h,decam_g,decam_r,decam_i,decam_z}` searches written for `Tile102008848…`; the GPU tiles of `342648` took 13:34–20:56 each (`_0,_2,_4-7` COMPLETED).
   The CPU route is the one the main experiment uses. Anomaly to read next: `342648_8` (`Tile102008532…`, GPU) was still RUNNING at 1:36 — five times its siblings. `342629_1/_3` still in `vis_lp` at 19:33. Next: pull, bundle with `output_sed`, magnitudes (µJy fluxes) vs `dr1_prelim_grade_ab_catalogue_csvs_20260623`; then tiles 1/3 SED on the CPU script.