prompt: pause pixelized-gradient-experiment (#100) for bedtime — resume=build samplers on kernel-CDF config

Author: Jammy2211

active.md

@@ -1,7 +1,14 @@
 # Active Tasks
 
 
-
-_No active tasks._
+## pixelized-gradient-experiment
+- issue: https://github.com/PyAutoLabs/autolens_workspace_developer/issues/100
+- status: in-progress, PAUSED for bedtime 2026-07-14 — feasibility ANSWERED (pix gradients DO work; certified test imaging_pixelization.py passed exit 0). RESUME = build the sampler experiment on the CORRECT config.
+- worktree: ~/Code/PyAutoLabs-wt/pixelized-gradient-experiment (autolens_workspace_developer on feature/pixelized-gradient-experiment; probes + pix_gradient_findings.md committed+pushed, NOT PR'd)
+- autonomy: supervised (research)
+- finding: can af.MultiStartAdam/ADABelief/Lion work for a pixelized source? YES — pix likelihoods are gradient-differentiable. My first probe's "no" was a methodology error (human caught it; certified in autolens_workspace_test/scripts/jax_grad/imaging_pixelization.py, re-run passed). Correct config = kernel-CDF mesh RectangularKernelAdaptDensity(bandwidth=0.1) [os_pix=1] OR adaptive mesh at over_sample_size_pixelization=4; truth-centred GaussianPriors; small FD step sweep near truth.
+- resume: (1) rewrite searches_minimal/probe_grad_pix.py to certified recipe (kernel-CDF + truth-centred priors + small-step FD near truth) → confirm OK; (2) build SLaM-pix-1 objective (fixed MGE lens light, free Isothermal+shear, kernel-CDF source) and run af.MultiStartAdam/ADABelief/Lion + af.Nautilus baseline locally; (3) A100 on RAL. REAL question = can multi-start gradient descent from BROAD starts recover the mass basin vs Nautilus (gradient correctness settled=yes). See [[project_pixelized_gradient_sampler_infeasible]] memory.
+- repos:
+  - autolens_workspace_developer: feature/pixelized-gradient-experiment
 
 

active/pixelized_gradient_sampler_experiment.md

@@ -0,0 +1,34 @@
+# Pixelized gradient-sampler experiment — can MultiStartAdam/ADABelief/Lion work for pix?
+
+Type: experiment
+Target: autolens_workspace_developer
+Repos:
+- @autolens_workspace_developer
+Difficulty: large
+Autonomy: supervised
+Priority: normal
+Status: issued
+
+Research question: do the newly-promoted multi-start gradient MAP optimizers
+(af.MultiStartAdam / MultiStartADABelief / MultiStartLion, Fit#1369) work on a
+PIXELIZED source reconstruction, not just the MGE likelihood the benchmark used?
+
+Setup (searches_minimal/, extending the MGE benchmark harness):
+- Model = SLaM SOURCE_PIX[1] style: lens MGE linear light with FIXED non-linear
+  geometry; lens mass (Isothermal + ExternalShear) FREE; source =
+  RectangularSplineAdaptImage (differentiable spline mesh) + adaptive
+  regularization (al.reg.Adapt); regularization coefficient FREE. Free non-linear
+  params ~= mass + shear + reg (~7-D). Adapt image bootstrapped from a quick
+  RectangularAdaptDensity+Constant fit (no adapt image needed), mirroring SLaM.
+- FD feasibility gate FIRST (probe_grad_pix.py): reverse-mode jax.grad of the
+  spline-pixelized log-evidence, FD-cross-checked. If FAIL_FD_MISMATCH, that IS
+  the answer — stop, report, no A100 burn.
+- Samplers: af.MultiStartAdam/ADABelief/Lion + af.Nautilus baseline.
+- Runs: local CPU smoke, then A100 on RAL (euclid_jump pipeline).
+- Deliverable: findings doc (do gradient MAP optimizers recover the mass basin
+  with a pixelized source vs Nautilus?).
+
+Decisions (human, 2026-07-14): SplineAdaptImage + adaptive reg; mass+reg free /
+light fixed. Checkpoint after the FD probe before samplers/A100.
+
+<!-- research spun off the multi-start gradient promotion; grad harness = searches_minimal -->