record: EP initializer-exception task shipped; fold the misfiled "EP nan" prompt - #130
Merged
Merged
Conversation
The 2026-08-03 nightly release leg failed `autofit_workspace_test scripts/graphical/ep.py` with the PyAutoFit initializer diagnostic, and was filed as an all-`nan` likelihood bug. It is not one. The exception lists "always returning `nan`" as a possible cause, but the guard that raises it is `np.allclose(fom[0], fom[1:])`, and `np.allclose` defaults to `equal_nan=False` — an all-`nan` figure-of-merit list returns False and cannot raise it. The detected condition is all-*equal* finite likelihoods, i.e. the degenerate factor state already characterised as defect 2 of PyAutoFit#1405. So this is that defect, not a fourth one. Folded into the follow-up-1 prompt rather than issued separately, carrying three things the release-leg run adds: - It is not hierarchical-specific. `ep.py` shares a `centre` prior across two `AnalysisFactor`s with no `HierarchicalFactor`, while `hierarchical.py` passed in the same shard — so scope the fix and its test to the per-factor update site in general. - It is release-profile-only: the smoke gate bypasses the sampler, so the initializer is never reached there. - The misleading `nan` line in the exception text should be corrected as part of the fix, at both raise sites (`samples_from_model` and `samples_jax`). Also carries forward the folded prompt's objection — that degrading the crash would turn the release leg green while EP still degenerates — as a loudness requirement rather than a reason to leave the crash in place. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GSBb1i58cNhTx64yxBGPfs
PyAutoFit#1454 merged (f02ea7e), so the prompt advances draft/ -> complete/ via lifecycle.py record. No separate issue was ever cut for it — the work sat under the PyAutoFit#1405 umbrella, which stays open for follow-up 2 (the scale-collapse basin, untouched). The record carries the traps rather than just the outcome, because most of the cost of this task was in them: the `nan` in the exception text is a red herring (np.allclose is False for nan, and nan draws are discarded before the check); `Status`'s third positional parameter is `updated`, not `flag`, so errored factor steps were logged to ep_history.csv as successes; a returned StatusFlag.FAILURE is routine and must not be counted; a consecutive-failure tally cannot catch the case it was written for, because zero KL between identical sweeps reads as convergence and ends the run first; and a warm output/ hides the whole bug by resuming past initialisation. Also records the posture change from #1405's "abort" to a loud warning that still returns, and the pre-existing HowToFit tutorial_5 bug found along the way and parked as NEEDS_FIX. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GSBb1i58cNhTx64yxBGPfs
Main gained a completion record (pr-ci-for-own-test-suite) after this branch was cut, so the index generated here was stale against the merge result even though it was fresh on the branch — which is what the drift check caught. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01GSBb1i58cNhTx64yxBGPfs
Jammy2211
pushed a commit
that referenced
this pull request
Aug 15, 2026
PyAutoBrain#130's stated close-condition named two remaining human decisions, and both turn out to be already answered: the scheduled-nightly standing grant was made 2026-07-09 (PyAutoBrain/AUTONOMY.md, predating the issue and live every night since), and the memory-system question is decided in the design doc's own §6 -- "memory earns its keep as consult-on-purpose context and fails reliably as guard-at-the-moment", with the conversion policy that follows from it. What was genuinely still open is a different §9 item: mitigation 6 shipped with a commitment to "review whether it went rote after ~10 ships", and nothing in the Mind tracked it -- no draft, active, planned, parked or ideas entry. Filing it here so closing #130 does not drop it. Filed through the Intake Agent (research / PyAutoBrain / small / safe); the body is expanded from intake's single-paragraph capture into the investigation sections, including the rubber-stamping signature to look for and a validate-the-instrument note, since a low firing rate and a broken lifter look the same from the outside (the D1 null-result-as-finding trap). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FTzN9AEF8DY2jTX2qxVHRk
Jammy2211
pushed a commit
that referenced
this pull request
Aug 18, 2026
PyAutoBrain#130's stated close-condition named two remaining human decisions, and both turn out to be already answered: the scheduled-nightly standing grant was made 2026-07-09 (PyAutoBrain/AUTONOMY.md, predating the issue and live every night since), and the memory-system question is decided in the design doc's own §6 -- "memory earns its keep as consult-on-purpose context and fails reliably as guard-at-the-moment", with the conversion policy that follows from it. What was genuinely still open is a different §9 item: mitigation 6 shipped with a commitment to "review whether it went rote after ~10 ships", and nothing in the Mind tracked it -- no draft, active, planned, parked or ideas entry. Filing it here so closing #130 does not drop it. Filed through the Intake Agent (research / PyAutoBrain / small / safe); the body is expanded from intake's single-paragraph capture into the investigation sections, including the rubber-stamping signature to look for and a validate-the-instrument note, since a low firing rate and a broken lifter look the same from the outside (the D1 null-result-as-finding trap). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FTzN9AEF8DY2jTX2qxVHRk
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Mind-side bookkeeping for PyAutoFit#1454 (merged,
f02ea7e).Two commits
1. Fold
graphical_ep_nan_likelihood_release_leg.mdinto the initializer prompt.The 2026-08-03 release-leg failure was filed as an all-
nanlikelihood bug. It is not one. The message it quoted is the body ofInitializerException, and the guard that raises it isnp.allcloseover the figures of merit — which isFalsefornan, andfigure_of_metricdiscardsnandraws before the check anyway. The condition detected is all-equal, finite likelihoods: PyAutoFit#1405's defect 2, on a non-hierarchical graph.Folded rather than issued as a fourth EP issue, carrying the three things the release-leg run added: it is not
HierarchicalFactor-specific (ep.pyshares a prior across twoAnalysisFactors and failed;hierarchical.pypassed in the same shard), it can only surface on the release leg (smoke bypasses the sampler), and the misleadingnanline should be corrected as part of the fix — which it was.2. Advance the prompt to
complete/2026/08/vialifecycle.py record.Its PR is merged. No separate issue existed — the work sat under the #1405 umbrella, which stays open for follow-up 2 (the scale-collapse basin, untouched).
On the record's shape
It leads with traps rather than outcome, because that is where this task's cost went:
nanred herring above, which sent the original triage after something that could not exist;Status's third positional parameter isupdated, notflag— so every errored factor step was written toep_history.csvas a success, and any "EP looks healthy" read off those CSVs pre-2026-08 is suspect;StatusFlag.FAILUREis routine (the Laplace optimiser emits one per failed line search) and must not be counted as a failure;output/hides the bug entirely by resuming past initialisation — a 12-run loop showed nothing until the directory was cleared each run, after which 2 of 8 runs hit it naturally.Also recorded: the posture change from #1405's "abort" to a loud warning that still returns, and the pre-existing HowToFit
tutorial_5bug found along the way (parked asNEEDS_FIX; PR on that repo's matching branch).lifecycle.py checkandlifecycle.py index --checkboth pass.Generated by Claude Code