Affected Repositories

• PyAutoFit (primary)
• autofit_workspace_test (regression script)
• PyAutoInsight (ledger), PyAutoCortex (criterion text; temp worktree off origin/main)

Branch Survey

Suggested branch: feature/ep-zero-success-factors — worktree ~/Code/PyAutoLabs-wt/ep-zero-success-factors/

Implementation Steps

1. Repro (worktree PyAutoFit on PYTHONPATH): cd /mnt/c/Users/Jammy/Science/active/analytic_gaussian && NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib python3 scripts/ensemble_seed.py --seed 140 --legs a --columns ep and seeds 29/38/71 --legs b; write to a scratch output dir so results/ens_n5/ (committed b44390d) is never overwritten. Read ep_history.csv / ep_diagnostics.results; capture the "..., iter=0" status messages. Add a projection="moments" leg (local patch of analytic_autofit.py if the script has no knob; nothing committed to the science repo).
2. Library: autofit/graphical/laplace/newton.py ~290–380 — a stop condition met at i == 0 with success=True is convergence: report updated=True, keep flag=NO_CHANGE. autofit/graphical/expectation_propagation/optimiser.py _track_failure / _stale_factor_warnings (~380–520) — a NO_CHANGE with an accepted projection counts as updated and its variables as seen; the per-(factor, variable) pass must not list a variable already at the fixed point. Check mean_field.py:511–620 (update_factor_mean_field, check_valid/update_invalid) so an accepted identical projection is not misclassified. Docstrings: optimiser.py NO_CHANGE paragraph, StatusFlag in utils.py.
3. Unit test in test_autofit/graphical/ (next to the feat(ep): report staleness per (factor, variable) and record reverted variables in ep_history.csv #1576 stale-tracking tests): a small Gaussian graph with one PriorFactor at its mode from the start → reported updated, no stale warning, history shows NO_CHANGE. Red before, green after.
4. Re-check seeds 29/71 leg B under projection="moments": HierarchicalFactor1 SUCCESS ≥ 1 expected (cured by feat: LaplaceOptimiser(projection="moments") — moment-matching projection for the hierarchical scatter (#1654) #1656). Record the outcome on this issue.
5. Workspace: autofit_workspace_test/scripts/graphical/analytic_gaussian_seed140.py (or a --seed leg of analytic_gaussian.py if it stays within the CI cap noted at line 62): leg A, seed 140, moments projection; asserts closed-form parity rows including widths, every factor SUCCESS ≥ 1 or accepted NO_CHANGE, STALE warning absent. Not in config/build/no_run.yaml.
6. Criterion 6: PyAutoCortex archive/tasks/analytic_gaussian/ensemble_parity.md row 6 + newest-first Log note in projects/analytic_gaussian.md (cortex.py check). Hand the human a diff for ensemble_seed.py:160 any_factor_never_succeeded.
7. Insight: tasks/ep_campaign.md phase-table 2026-10-23 entries + findings follow-up rows; campaigns.yaml graphical-ep.next/updated.

Key Files

• autofit/graphical/laplace/newton.py — zero-iteration convergence → updated semantics
• autofit/graphical/expectation_propagation/optimiser.py — stale tracking (_factors_skipped, _stale_factor_warnings)
• autofit/graphical/mean_field.py — projection acceptance
• autofit/graphical/utils.py — StatusFlag
• autofit_workspace_test/scripts/graphical/analytic_gaussian.py — pattern for the regression
• PyAutoInsight/tasks/ep_campaign.md, campaigns.yaml — ledger

Witness

python3 scripts/ensemble_seed.py --seed 140 --legs a --columns ep in analytic_gaussian writes a sidecar whose legs.A.info["ep stale"] is False and every factor has SUCCESS ≥ 1 — or, if the zero-SUCCESS factor is ruled legitimate, criterion 6 is amended with the reason; plus a regression in autofit_workspace_test/scripts/graphical/ pinning the leg-A case.

Lineage

PyAutoFit#1405, #1562, #1574, #1576, #1580, #1656; PyAutoCortex ruling R-20260910-04; Mind prompt draft/bug/autofit/ep_factors_end_the_run_with_zero.md; Insight tasks/ep_campaign.md.

EP factors end the run with zero SUCCESS updates on the exactly-Gaussian leg A of analytic_gaussian (seed 140) and on 18/200 leg-B seeds

Type: bug
Target: PyAutoFit
Repos:

• PyAutoFit
• autofit_workspace_test
  Themes:
• graphical-ep
  Difficulty: medium
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: glance
  Witness: python3 scripts/ensemble_seed.py --seed 140 --legs a --columns ep in analytic_gaussian writes a sidecar whose legs.A.info["ep stale"] is False and in whose legs.A.info["ep flags"] every factor has SUCCESS >= 1 — or, if a zero-SUCCESS factor is ruled legitimate, criterion 6 is amended with the reason; plus a regression in autofit_workspace_test/scripts/graphical/ that pins the leg-A case.
  Review-minutes: 3
  Unattended: ready
  Epic: graphical-ep
  Filed: 2026-09-10

Why

Criterion 6 of the pre-registered witness of the science task
PyAutoCortex/tasks/analytic_gaussian/ensemble_parity.md ("no seed in which any factor ends the
run with zero SUCCESS updates") missed: 19 of 400 seed-legs recorded at least one factor that
finished the EP run having never once been updated successfully — the STALE state PyAutoFit
#1562 / #1574 / #1576 / #1580 taught the library to report.

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

Affected seeds, from the 200-seed N=5 ensemble (RAL job 342413_[0-199], autofit 2026.8.17.1):

• leg B (sigma ~ TruncatedGaussian(10, 5, 0, 100)): seeds 29, 38, 44, 52, 62, 71, 75, 83,
  90, 106, 133, 138, 147, 149, 152, 160, 178, 193
• leg A (sigma = 10 fixed): seed 140

Leg A is the damning one. It is a linear-Gaussian graph — mu ~ N(50, 10^2),
x_i | mu ~ N(mu, 10^2), y_ij | x_i ~ N(x_i, s_i^2) — on which Laplace EP is exact by
construction, and criterion 3 confirms it: autofit EP recovers leg A 200/200 with worst
a = 0.001. A factor that never takes an update on a graph where every projection is available
in closed form is a library defect, not a model limitation.

The sidecars show two distinct shapes:

• a PriorFactor* that ends with only NO_CHANGE (leg A seed 140: PriorFactor5 {NO_CHANGE: 2};
  most leg-B seeds, e.g. 38/44/52/62 PriorFactor17 {NO_CHANGE: 3})
• a HierarchicalFactor1 that ends with only BAD_PROJECTION and never a SUCCESS
  (leg B seeds 29, 71, 106, 152, 178, 193 — e.g. {BAD_PROJECTION: 15}, {BAD_PROJECTION: 20})

Whether those are one defect or two is part of the question.

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 140 --legs a --columns ep

and read the EP flag fields of the sidecar it writes — in
results/ens_n5/seed_0140.json they are legs.A.info["ep flags"]
(a {factor_name: {FLAG: count}} map) and the derived boolean legs.A.info["ep stale"].
The witness scripts are borrowed verbatim from
autofit_workspace_test/scripts/graphical/analytic_{reference,ep_minimal,autofit}.py @ 165a19e.

Ask

• Which factor is it, on which update, and why does it never succeed — is the projection
  rejected, is the update a no-op because the message is already the cavity, or is the factor
  never visited?
• Are the NO_CHANGE-only PriorFactor case and the BAD_PROJECTION-only HierarchicalFactor
  case the same defect or two?
• Fix it, or explain why a zero-SUCCESS factor is legitimate on an exactly-Gaussian graph — and
  if it is legitimate, criterion 6 is what has to change, with the reason recorded.
• Add a regression to autofit_workspace_test/scripts/graphical/ that pins the leg-A seed-140
  case: on the exactly-Gaussian leg every factor ends with at least one SUCCESS update.

Lineage

• PyAutoFit#1405 — still open, the Laplace-on-scatter collapse basin
• PyAutoFit#1580 — merged, the most recent STALE-reporting fix; this evidence is from a mirror
  that contains it (RAL PyAutoFit 66f9f8d5d contains 08207bad0)
• PyAutoCortex/rulings/2026/09/R-20260910-04.md — the ruling
• 2026-09-30: LaplaceOptimiser(projection="moments") (PyAutoFit#1656, record complete/2026/09/ep-moment-projection.md) gives 3/3 HierarchicalFactor SUCCESS on the tight toy (50, 50.5, 49.5) where the default mode path gives 0 (all BAD_PROJECTION); the BAD_PROJECTION-only HierarchicalFactor case may be cured under projection="moments" — re-check seed 140 / the 18 leg-B seeds with it before fixing the mode path