Affected Repositories

• PyAutoFit (primary)
• autofit_workspace_test (regression)

Branch Survey

Suggested branch: feature/ep-evidence-nonfinite — worktree ~/Code/PyAutoLabs-wt/ep-evidence-nonfinite/

Implementation Steps

1. Repro: scratch copy of scripts/ + config/; python3 -W error::RuntimeWarning scripts/ensemble_seed.py --seed 25 --columns ep (and 172) with the worktree PyAutoFit on PYTHONPATH, NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib; then a normal run with EPMeanField.log_evidence monkeypatched to log each term (variable name, message mean/variance/limits, factor log_norm). Never write into the science checkout's results/.
2. Library (autofit/graphical/expectation_propagation/): ep_mean_field.py log_evidence (~lines 310–336) — per-term dicts, np.isfinite scan, one logger.warning per instance naming variables/factors, np.errstate(invalid="ignore") around the sum, explicit float("nan"); diagnostics.py (~131) — record evidence_nonfinite: <names> in ep_diagnostics.results; history.py evidence_divergence (~152) — NaN → inf.
3. Tests in test_autofit/graphical/functionality/ (new test_evidence_nonfinite.py or in test_history.py): build an EPMeanField from a tiny graph, inject a non-finite log_norm / degenerate message; assert no RuntimeWarning, one named warning, NaN evidence, inf divergence. pytest test_autofit/graphical -q -x, then the full suite.
4. Workspace: scripts/graphical/analytic_gaussian.py — warnings.catch_warnings(record=True) around the EP column, assert no RuntimeWarning whose filename ends in ep_mean_field.py, assert np.isfinite on the final log_evidence; seed 25 leg only if ≤ 60 s.
5. Ledgers: append a 2026-10-23 row to the findings follow-ups table in PyAutoInsight tasks/ep_campaign_state_2026-10-23.md (render with the worktree's own bin/pyauto-insight board + PYAUTO_BRAIN), and a Log note in PyAutoCortex projects/analytic_gaussian.md.

Key Files

• autofit/graphical/expectation_propagation/ep_mean_field.py — variable_evidence, factor_evidence, log_evidence
• autofit/messages/abstract.py:380 — log_normalisation (read only)
• autofit/graphical/expectation_propagation/history.py — evidence_divergence, evidences
• autofit/graphical/expectation_propagation/diagnostics.py — results file
• autofit_workspace_test/scripts/graphical/analytic_gaussian.py — regression

Witness

python3 scripts/ensemble_seed.py --seed 25 --columns ep and the same for seed 172 complete with no RuntimeWarning from ep_mean_field.py, and the EP evidence they record is finite — or the non-finite term is raised/logged with the variable named rather than warned-and-summed, with a regression in autofit_workspace_test/scripts/graphical/ asserting the EP evidence of a converged fit is finite.

Lineage

PyAutoCortex ruling R-20260910-04; PyAutoFit#1405; PyAutoFit#1711/#1712 (sibling finding); Mind prompt draft/bug/autofit/nan_in_ep_evidence_accumulation_invalid_value.md.

NaN in EP evidence accumulation: invalid value encountered in add in ep_mean_field.py on analytic_gaussian seeds 25 and 172

Type: bug
Target: PyAutoFit
Repos:

• PyAutoFit
• autofit_workspace_test
  Themes:
• graphical-ep
  Difficulty: medium
  Autonomy: supervised
  Priority: medium
  Status: formalised
  Consequence: glance
  Witness: python3 scripts/ensemble_seed.py --seed 25 --columns ep and the same for seed 172 complete with no RuntimeWarning from ep_mean_field.py, and the EP evidence they record is finite — or the non-finite term is raised/logged with the variable named rather than warned-and-summed, with a regression in autofit_workspace_test/scripts/graphical/ asserting the EP evidence of a converged fit is finite.
  Review-minutes: 3
  Unattended: ready
  Epic: graphical-ep
  Filed: 2026-09-10

Why

Two seeds of the 200-seed N=5 ensemble behind
PyAutoCortex/tasks/analytic_gaussian/ensemble_parity.md (RAL job 342413_[0-199],
autofit 2026.8.17.1) emitted a numpy warning from the EP evidence accumulator, repeatedly.
Verbatim from hpc/batch_cpu/error/error.342413_25.err (line 1, repeated 7x in that file and
10x in error.342413_172.err):

/mnt/ral/jnightin/PyAuto/PyAutoFit/autofit/graphical/expectation_propagation/ep_mean_field.py:336: RuntimeWarning: invalid value encountered in add
  return factor_evidence + sum(variable_evidence.values())

invalid value encountered in add is numpy's NaN-producing-add: at least one term of
factor_evidence or of variable_evidence.values() is non-finite (NaN, or a ±inf pair), and
the library warns and keeps going, folding the result into the EP evidence readout.

This is a separate symptom from the zero-SUCCESS stale-factor finding of the same ruling:
neither seed 25 nor seed 172 appears in that finding's 19 stale seed-legs, and both seeds'
parity results are otherwise fine — they are not among the outliers on any row. So the evidence
accumulator is going non-finite on runs that otherwise look healthy, which is exactly the case a
warning is least likely to be noticed in.

Ruled 2026-09-10 as PyAutoCortex/rulings/2026/09/R-20260910-04.md (accept with findings).

Repro

Science project analytic_gaussian at /mnt/c/Users/Jammy/Science/analytic_gaussian
(git main, no remote), results committed at b44390d:

cd /mnt/c/Users/Jammy/Science/analytic_gaussian
python3 scripts/ensemble_seed.py --seed 25 --columns ep
python3 scripts/ensemble_seed.py --seed 172 --columns ep

Run them with python3 -W error::RuntimeWarning (or np.seterr(invalid="raise")) to get the
traceback at the moment the term goes non-finite. The witness scripts are borrowed verbatim from
autofit_workspace_test/scripts/graphical/analytic_{reference,ep_minimal,autofit}.py @ 165a19e;
the per-seed sidecars are results/ens_n5/seed_0025.json and seed_0172.json.

Ask

• Where does the non-finite variable evidence come from? Which variable's
  variable_evidence entry is NaN/inf, and what makes it so — a degenerate (zero- or
  negative-variance) message, a log of a non-positive normalisation, an inf−inf cancellation?
• Should this warn and continue at all? Decide between guarding it (skip / clamp the term
  with a named reason), raising it (a non-finite evidence is a broken fit and should say so), or
  logging it with the offending variable named. Warn-and-sum is the one option that leaves no
  usable trace in a batch run — the only reason it was caught here is that SLURM kept the
  .err files.
• Does it bias the EP evidence readout? If the accumulated evidence is used for
  convergence, model comparison or the reported log_evidence, a silently-NaN term is a
  correctness problem, not cosmetic. Say which, with the evidence for it.
• Add the regression named in the Witness above.

Lineage

• PyAutoCortex/rulings/2026/09/R-20260910-04.md — the ruling that surfaced it
• PyAutoFit#1405 — the open Laplace-on-scatter collapse basin (leg B's known bias); this is not
  that, but the same EP path