[Jammy2211]

Summary

A research note answering autolens_profiling#515: did the two gradient defects found in PyAutoArray#671 affect the phase-6 gradient campaign? No. None of the 73 results/**/gradient.json verdicts changes, the day_zero_gradient snapshot stands, and no cell needs a re-run on account of #671.

• Every pixelized campaign cell is --standard-size and is evaluated at iid draw 0 of the seeded production sequence, not at the prior medians. The mass centre there is a generic value near 0, never exactly 0, so neither defect's condition is met:
  • no exact ± border ties, so the border-relocator max tie gradient (fixed in PyAutoArray#672) cannot trigger;
  • no grid point on the mass centre, so the Isothermal 1e-8 core cannot come into play.
• The cells really evaluated at the medians are not pixelized. The imaging MGE cells also hold the mass fixed.
• Measured control: interferometer/rectangular/sma, re-run with the campaign flags, gives a bit-identical gradient before and after #672 (largest relative change 2.4e-16).
• The remaining autodiff-vs-FD disagreements are listed by family. None is from #671: FD noise on small gradients, an unexplained CPU-sparse vs GPU-dense gap in four Delaunay cells, an unexplained 0.1–0.4 % imaging-MGE gap, and the known zero gradient of image-plane cells.
• Side finding: every standard driver except group/multi-galaxy records "point": "prior_medians" while evaluating at iid draw 0.

Scripts Changed

• None. Documentation only:
  • wiki/research/gradient_campaign_impact_border_tie.md (new research note);
  • wiki/README.md (link to it under research/).

Test Plan

• check_wiki.py --check, build_readme.py --check and check_results_layout.py --check pass locally.
• No scripts changed, so no smoke scope. Analysis scripts live in Science/scratch/gradient_campaign_impact_2026-11-02/.

Closes #515 together with PyAutoLabs/PyAutoPulse#89 (knowledge page).

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code