Overview

fix: guard MCMC thinning and short-chain autocorrelation

Plan

Clamp Emcee/Zeus thinning to at least one (no thinning for autocorrelation time below two). Treat unavailable previous autocorrelation estimates as None with a diagnostic, without masking unrelated errors. Remove short-chain check_size workarounds from alignment regressions and test real short Emcee runs plus aligned thin=1 samples.

Tier: glance — merge mode: human /prm; this session stops at PR-open by explicit agreement.

Detailed implementation plan

Affected repository: PyAutoLabs/PyAutoFit. Files: autofit/non_linear/search/mcmc/{emcee,zeus}/search.py; test_autofit/non_linear/search/mcmc/test_{emcee,zeus}.py.
Clamp Emcee/Zeus thinning to at least one (no thinning for autocorrelation time below two). Treat unavailable previous autocorrelation estimates as None with a diagnostic, without masking unrelated errors. Remove short-chain check_size workarounds from alignment regressions and test real short Emcee runs plus aligned thin=1 samples.
Run focused regression tests and the full PyAutoFit suite; compare failures against main rather than claiming unverified green. No public signature changes planned.

Branch survey: main clean; no existing PyAutoFit claim in current Mind active.md; no open PyAutoFit PRs or matching open issue found. Branch: feature/mcmc-short-chain-guards. These four separately reviewed tasks are coordinated in one user-authorized group, with isolated worktrees and sequential edits.

Authorization and Heart RED

User request: "Find 4 grouped tasks and do them". After the exact Heart RED reasons were disclosed and development-only authorization requested, the user replied "Yes I authorise". Applies to selecting, implementing and testing these four tasks and opening reviewable PRs; no merge or release. Tests and author diff review will be recorded at PR-open, not claimed in advance.

Published RED reasons:

• install verification FAILED (testpypi; checks F)
• release validation FAILED (stage integrate)
• PyAutoHeart/workflows/Release Integrate: failure
• manifest_drift/checks/workspace checkouts (manifest ↔ disk): red
• Release readiness: release validation FAILED (stage integrate)
• Release readiness: install verification FAILED (testpypi; checks F)
• PyAutoArray/test_autoarray/inversion/inversion/interferometer/test_interferometer.py::test__interferometer_sparse_operator__mapped_reconstructed_data_dict_matc…
• Unit-test timing: 1 test regressions (>3× baseline), 11 slow (>1.5×)
• Release validation: NOT release_ready — v2026.10.25.1.dev83701 profile=release (2026-10-25T07:21:05+00:00)
• validation_report/stages/integrate: fail
• verify_install FAILED: verify_install FAILED
• F: colab gate: finufft (declared by autoarray, absent on Colab, imported)
• Install verification: FAILED (testpypi; F) (2026-10-25T07:42:48.134987+00:00)
• verify_install: red
• PyAutoBrain: red
• PyAutoLens: red
• autolens_profiling: red
• Worktree drift: 1 orphan / 0 missing / 3 dirty

Original prompt

Emcee/Zeus: thin can be 0, and check_size blows up on short chains

Type: bug
Target: PyAutoFit
Repos:

• PyAutoFit
  Themes:
• mcmc
• test-mode
• robustness
  Difficulty: small
  Autonomy: supervised
  Priority: normal
  Status: formalised
  Consequence: glance
  Witness: thin is max(1, int(max(times) / 2)) in both emcee/search.py and zeus/search.py, pinned by a unit test on a chain whose largest auto-correlation time is below 2 (thin == 1, no ValueError); Emcee guards auto_correlations_from on chains shorter than check_size the way Zeus does, so af.Emcee(nsteps=10) under PYAUTO_TEST_MODE=1 completes; and the fix: Emcee/Zeus pair thinned parameters with unthinned log-probabilities #1628 regression tests drop their AutoCorrelationsSettings(check_size=5, ...) workaround.
  Review-minutes: 3
  Unattended: ready
  Filed: 2026-09-14
  Parent: draft/bug/autofit/emcee_zeus_samples_log_prob_misalignment.md

Two pre-existing robustness defects in the MCMC searches, found while fixing
PyAutoFit#1628 and deliberately left out of that PR because both change values
or behaviour rather than correcting a misalignment.

1. thin can be 0, which raises

Both emcee/search.py and zeus/search.py compute:

thin = int(np.max(auto_correlations.times) / 2.0)

This is 0 whenever the largest auto-correlation time is below 2.0 — a short or
well-mixed chain. A zero slice step then raises
ValueError: slice step cannot be zero inside get_chain, and (after #1628)
inside get_log_prob as well.

The fix is presumably max(1, ...), but that changes the thin value on
affected runs, so it needs a deliberate decision rather than a silent clamp:
confirm thin=1 (no thinning) is the intended behaviour for a chain whose
auto-correlation time is under 2, rather than an error worth surfacing.

2. check_size=100 blows up on chains shorter than 100

auto_correlations_from slices chain[:-check_size]. With the default
check_size=100 and a chain shorter than that, the slice is empty and
emcee.autocorr raises IndexError before the samples object is ever built.

This is always true under PYAUTO_TEST_MODE=1, where nsteps=10.

Zeus already guards its equivalent with except IndexError; Emcee does not,
so the two searches behave differently on the same input. Either Emcee should
gain the same guard, or both should fail with a message naming check_size and
the chain length instead of an opaque IndexError from a third-party module.

Worked around in the #1628 regression tests by passing
AutoCorrelationsSettings(check_for_convergence=False, check_size=5, required_length=2)
— that workaround should become unnecessary once this is fixed.

Note

Neither is a correctness bug in results: they are crashes and an inconsistency
between the two searches. Do not bundle them with a results-changing fix.

Folded 2026-09-17

draft/bug/autofit/emcee_crashes_in_autocorrelation_when_the_chain.md (filed
2026-09-10) described item 2 above and was removed at the witness-campaign
close-out (PyAutoMind#398). Its witness — af.Emcee(nwalkers=10, nsteps=50).fit(...)
on the 1D Gaussian example completes with a warning instead of raising
IndexError — is part of this prompt's witness now (Emcee gains Zeus's guard).

Implementation and validation — 2026-10-25

PR: #1730

Full library suite: 3922 passed, 1 skipped, 1 xfailed (Python 3.12.14). Focused MCMC tests: 14 passed, including real Emcee/Zeus reduced fits with default check_size=100. Additional real Emcee smoke: 10 walkers, 50 steps, test mode disabled, samples built and previous_times=None. Six new regression cases fail on unchanged main (zero thin and empty previous chain), while ordinary thinning remains unchanged. Author diff review passed.

Published vitals: organism RED / PyAutoFit scope GREEN. Development-only authorization applies; no merge or release. Awaiting required CI and human review.