Overview

Make start-point plotting tests deterministic. Selected as one of four PyAutoFit model-inspection and test-reliability tasks authorized on 2026-10-25.

Plan

• Reproduce the reported defect against current main.
• Set an explicit seed in the shared Emcee test factory using the current search seed interface. Repeat all three tests 200 times; leave production initializer tolerance unchanged.
• Run focused regressions and the full library suite; record any environmental or pre-existing failures honestly.

Tier: glance (notify for test-only task) — merge mode: human /prm; no merge requested.

Detailed implementation plan

Set an explicit seed in the shared Emcee test factory using the current search seed interface. Repeat all three tests 200 times; leave production initializer tolerance unchanged.

Repository: PyAutoFit; main clean at discovery; no existing active PyAutoFit claim or open PR found. Branch: feature/seed-start-point-tests. Each task retains a separate branch and PR. The four branches are coordinated within this user-authorized group, with no overlapping source edits committed together. Existing identifiers/output-directory identity must remain stable.

Original Prompt

test_start_points.py draws unseeded start points and flakes on IDENTICAL_FIGURES_OF_MERIT (nojax leg, PyAutoFit#1714)

Type: bug
Target: PyAutoFit
Repos:

• PyAutoFit
  Themes:
• ci
• samplers
  Difficulty: small
  Autonomy: safe
  Priority: normal
  Status: formalised
  Consequence: notify
  Witness: pytest test_autofit/non_linear/search/test_start_points.py -q -p no:randomly --count 200 (or a 200-iteration shell loop) passes every time; the three tests draw their points from a seeded generator (or a fitness whose figures of merit cannot coincide), and no InitializerException can arise from the draw.
  Review-minutes: 2
  Unattended: ready
  Filed: 2026-10-23

Why

On PyAutoFit PR #1714 (run 39976190947, unittest-nojax leg) test_plot_false_skips_the_plot
failed with autofit.exc.InitializerException: The initial samples all have the same figure of merit while the 3.12 and 3.13 legs of the same sha passed and main's nojax leg was green an hour
earlier (run 39970496109). The PR touches only the EP evidence readout; the test does not reach it.

_search_and_fitness builds af.Emcee() with af.Model(af.ex.Gaussian) on data = np.ones(10)
and calls search.start_points(model, fitness, n=2). InitializerBall.samples_from_model draws
from np.random.default_rng() when no rng is passed (autofit/non_linear/initializer.py:104) and
raises when np.allclose(fom[0], fom[1:]) (:214). With n=2 on flat data two random draws land
on near-identical log likelihoods often enough to flake. Locally the test passed 3/3 under a nojax
emulation on the PR head.

Ask

• Seed the draws in these tests (pass rng=np.random.default_rng(0) through start_points, or set
  the search's seed) so the tests are deterministic, or use a fitness whose figures of merit cannot
  coincide for distinct points.
• Check the sibling tests in the file (test_plots_the_first_point_once,
  test_initializer_gets_the_effective_core_count) for the same unseeded draw.
• Consider whether IDENTICAL_FIGURES_OF_MERIT at n == 2 should use a tolerance tighter than
  np.allclose defaults, or require n >= 3, in the library itself — report, do not change, unless
  trivially safe.

Lineage

• PyAutoFit#1714 — the PR whose nojax leg flaked (log excerpt in the issue bug: EP log_evidence sums a non-finite term silently (NaN evidence on analytic_gaussian seeds 25/172) #1713 comment of 2026-10-23)
• search-extensibility A2 — the start_points entry point the test covers