Validation: PJ011646 WFC3-IR reduction vs Aris's dataset

[Jammy2211]

Overview

Validate PyAutoReduce's WFC3-IR path on a science-grade lens: reduce PJ011646 (Planck/PASSAGES lensed dusty starburst, z_lens 0.555, z_source 2.125) from archival WFC3 data and compare the modeling-ready product against Aris's dataset at /mnt/c/Users/Jammy/Science/aris_PJ011646/dataset/aris/PJ011646 (201×201 @ 0.0642″/pix, 3.8″ mask). Second lens of the reduction-validation series — slacs1430+4105 (#17) validated the ACS path and settled the comparison methodology; this task exercises the genuinely-different IR deltas (_flt inputs, up-the-ramp CR handling, half-native drizzle scale, fine-grid Casertano R).

Plan

• Inspect Aris's dataset (real-data gate): plot data/noise/PSF; the bundle ships no positions, extra-galaxy masks or coordinates — create positions + field-object notes at inspection; resolve PJ011646 coordinates + HST program/filter via MAST and literature (filter confirmed empirically by PSF FWHM if ambiguous).
• scripts/reduce_pj011646.py: instrument="wfc3_ir", final_scale=0.0642, 201×201 cutout matching Aris's grid; respect the J0252 IR pixfrac/zero-weight-holes finding (docs/design/wfc3.md).
• Pixel parity with the slacs1430 lessons: dihedral × rotation orientation scan over source structure FIRST; lens-anchored registration; ratio maps + S/N; explicit does-the-legacy-noise-carry-R check (IR R is materially larger on the fine-grid branch); PSF radial profiles.
• Model parity: identical MGE + SIE + shear + MGE fits (Nautilus), 3.8″ mask, shared positions through the measured anchor transform; fits run SERIALLY on a quiet machine (~5.5 GB each, slacs1430 OOM lesson).
• Verdict here; pipeline defects escalate as PyAutoMind prompts (PyAutoReduce source claimed by jwst-frame-feasibility; no source edits in this task).

Detailed implementation plan

Work classification

Workspace (scripts + analysis; no library source edits). Autonomy: safe effective (--auto, plan human-approved in-session 2026-07-10). Brain suggested re-homing as research/; human classified test at intake — kept.

Affected repositories

• PyAutoReduce (primary — new scripts/reduce_pj011646.py + parity-fit prototype)
• autolens_assistant (driver only; no commits; scratch under scripts/scratch/pj011646_parity/)

Branch survey (2026-07-10)

Repository	Current Branch	Dirty?
./PyAutoReduce	main	2 untracked (keck-ao parked analysis + gitignored output/)
./autolens_assistant	feature/assistant-ref-mechanics (other task, in-place)	driver-only here, no commits

Suggested branch: feature/pj011646-wfc3-parity (created only at ship). Claim handling: PyAutoReduce worktree held by jwst-frame-feasibility → keck-ao/slacs1430 pattern: analysis on the main checkout, no claim, outputs to gitignored scripts/output/; script commit + PR gated until the claim releases (or human-directed).

Key files

• PyAutoReduce/scripts/reduce_slacs1430.py, prototypes/slacs1430_parity_fit.py — the shipped templates (0face12)
• PyAutoReduce/docs/design/wfc3.md — IR deltas: _flt, 0.128″ native → 0.065″ recommended, pixfrac/coverage finding, R on the fine-grid branch
• autolens_assistant/scripts/scratch/slacs1430_parity/pixel_comparison.py — dihedral-aware comparison machinery to adapt
• /mnt/c/Users/Jammy/Science/aris_PJ011646/dataset/aris/PJ011646/ — data (201×201 f4), noise (f8), psf (51×51), info.json only
• PyAutoReduce#17 — methodology notes this task inherits

Original Prompt

Click to expand starting prompt

PyAutoReduce validation: PJ011646 WFC3 reduction vs Aris's dataset

Type: test
Target: workspaces
Repos:

• autolens_assistant
• PyAutoReduce
  Difficulty: medium
  Autonomy: safe
  Priority: normal
  Status: formalised

Follow-up to the slacs1430+4105 PyAutoReduce validation: reduce the lens PJ011646 from archival HST/WFC3 data using PyAutoReduce and compare the result to Aris's existing dataset at /mnt/c/Users/Jammy/Science/aris_PJ011646/dataset/aris/PJ011646. Same comparison methodology as the SLACS task (image/noise/PSF residuals plus lens-model parity fits via autolens_assistant); exercises the WFC3 reduction path rather than ACS. Blocked-by: the slacs1430+4105 comparison task should ship first so the methodology is settled.

[Jammy2211]

Phase 0 complete — target resolved, Aris dataset inspected (real-data gate).

• Target: MAST resolves PJ011646 as J011646.77 at (19.194875°, −24.617194°), program 14653 (PASSAGES snapshot), WFC3/IR F160W, ~2596 s total — matches Aris's 0.0642″/pix (half-native IR) drizzle.
• The lens: large Einstein ring, radius ≈ 2.2–2.8″, multiple bright arc knots (S/N 37–64), bright central lens galaxy; peak S/N 134; no NaNs/holes. 201×201 @ 0.0642″ = 12.9″ FoV; 3.8″ fit mask per info.json.
• Field objects (all outside the mask): bright companion at (−1.0″, −5.2″) r=5.3″, compact galaxy at (+3.0″, −3.2″) r=4.4″, corner object at (+4.1″, +6.4″); a few faint smudges S. Gate flag (non-blocking): nothing inside the mask looks artefactual; if any of these deserve extra-galaxy treatment inside 3.8″, say so here.
• Created positions (bundle ships none): (+2.05, +0.58) and (−2.31, +0.19) — the N/S knot pair at matching radii (2.1″/2.3″), used identically on both datasets with threshold 0.7″; if the pair proves non-conjugate the penalty drops from both fits (differential parity unaffected).

Inspection plots: autolens_assistant/scripts/scratch/pj011646_parity/plots/aris_dataset.png, aris_data_annotated.png. Next: reduce_pj011646.py via the wfc3_ir adapter at Aris's exact grid.

[Jammy2211]

Phase 1, first pass — the packaging guard caught a real IR gap; re-run in flight.

The pipeline ran MAST → CRDS → AstroDrizzle (5 direct _flt exposures — the merged is_direct_product hygiene worked, no HAP duplicates) but failed loudly at packaging: a single 123-px structured zero-coverage hole in the cutout. Diagnosis:

• The hole sits at (−4.0″, −3.4″), r ≈ 5.3″ — outside the 3.8″ science mask (0 bad px inside).
• All five exposures carry DQ 512 (WFC3-IR blob) at the same detector pixels, and the snapshot dithers are only 2–6 px, so the blob never moves off that sky spot: ctx = 0, wht = 0 — no valid input exists there.
• Root cause: the pipeline has no DQ-bits dial — drizzle_kwargs_for never sets final_bits, so AstroDrizzle's default rejects every flagged pixel. Standard IR practice passes 512 as usable; Aris's reduction of the same data has no hole.

Filed as PyAutoMind/bug/pyautoreduce/wfc3_ir_dq_bits_dial.md (adapter-level final_bits with an IR default including 512 + design-doc documentation + a question about whether the guard should distinguish defects inside vs outside the science aperture). No source edits here — repo claimed by jwst-frame-feasibility.

Workaround for this validation (documented in reduce_pj011646.py): inject final_bits=driz_sep_bits=512, matching standard practice and Aris's treatment. Clean re-run is in flight. Two validations, two loud-guard catches — the guard is earning its keep.

[Jammy2211]

Phase 2 complete — pixel parity PASS on data; the noise maps disagree for understood, quantified reasons (both sides implicated).

Orientation: Aris's frame is the pipeline frame rotated −71.1° (no reflection) — found by the upgraded full-range (±180°) dihedral × rotation scan, ring correlation 0.997 (same quality as slacs1430). Lesson from #17 upgraded: never trust a narrow scan window or an eyeball on a near-circular ring.

Quantity	Result	Reading
data ratio (bright px, n=12025)	1.016 [0.71, 1.39]	photometric parity; spread wider than slacs1430 (large-angle interpolation + PSF-width mismatch + knotty structure)
empirical blank-sky scatter	aris 0.0090, ours 0.0077 e−/s	underlying data quality comparable (ours slightly smoother — pixfrac 1.0)
noise-map ratio	≈4.9 uniform	decomposes as R(2.39) × our pre-R conservatism (1.58) × Aris under-prediction (~1.3)
Aris noise vs his own sky scatter	0.62	his map under-predicts the data's own blank-sky scatter by ~1.6× and carries no R → overstated S/N in past modeling of this dataset
our pre-R noise vs empirical	1.58	IR recipe conservatism — flagged for the wfc3_ir_dq_bits_dial prompt pickup to examine alongside the bits work
PSF FWHM	ours 0.259″ vs aris 0.202″	tier-1 ePSF, 71 stars; F160W diffraction ~0.19″ — the gap needs a look at Aris's PSF provenance (may be TinyTim/native)
residual sky	ours 0.0169 vs aris 0.0096 e−/s	same family as slacs1430/slacs0008 sky-subtraction offsets

Figures: autolens_assistant/scripts/scratch/pj011646_parity/plots/{pixel_comparison,psf_comparison,orientation_check}.png.

Phase 3 next (serial Nautilus fits, identical model both datasets). Expectation to test: parameter medians agree; χ²/posterior widths differ per the noise table above — which is exactly the "does the science change" question this task exists to answer.