Status: resolved (2026-11-13). No wrong result, crash or missing check. Mistral’s four probes compared the batched messages, beliefs and free energies of tied cluster nodes over a plate with the per-replica ones (several factors per cluster, several precision classes, Gaussian inputs with points, discrete and continuous targets, pruning categorical inputs, N = 2, scalar points, flat directions, rows whose factors differ) and the errors of both paths; all agree. The review wrote no answer (it stopped at its turn limit); its probes, after fixes to their own code, are the evidence. Two became regression tests. Related: Feature Backlog (I19), LP-032 Findings of the review of I6.
Sources: a review of the uncommitted I19 change by Mistral (
vibe --agent review, 100 turns, stopped at the limit without an answer file), run with Julia in the helper worktreeLenticulum.jl-mistralat1770cc13with the diff applied; probes inscratch/i19/of that worktree, copied with the corrected ones (-fixed), the brief and the run’s output to../Lenticulum.jl-Review-2026-11-13/; the logs of the corrected runs in~/.cache/lenticulum/i19/probes/. code:lib/Mycelium.jl/src/cluster_batch.jl(_batch_joint,_as_clusters,_same_struct),lib/Mycelium.jl/src/cluster.jl(_continuous_parts,_discrete_parts,_belief_parts); tests:lib/Lenticulum.jl/test/regime_plate.jl,test/cluster_flat.jl,test/cluster.jl.
Problem. I19 replaced the per-replica joints of batched cluster nodes over a plate by one
stacked joint whose data terms are computed for all replicas together, and assembled each
replica’s message and belief from its columns (cluster.md, “Batching in one stacked joint”).
The requirement: for every row _cluster_batchable accepts, the results equal the per-replica
path’s up to rounding, and the same errors are thrown.
| # | probe | result | what was done | test |
|---|---|---|---|---|
| P1 | multi-factor clusters (a sum constraint and a RegimeFactor) over a plate: messages to every channel, beliefs, graded free energies, against per replica; N = 2 and 5, scalar and vector points, Gaussian inputs with points | 1 188 checks pass | — | covered by cluster.jl and regime_plate.jl (random instances) |
| P2 | rows of RegimeFactors: N = 2, pruning categorical inputs, observed switches (not batchable), a row with one unequal factor, _as_clusters sharing one cluster | 906 checks pass after two fixes to the probe (a tolerance written atol = 1e-9 && …, which Julia parses as the tolerance; a generic comparison that read a Dict’s unset slots and threw UndefRefError) | the unequal-factor part adopted | ”RegimeFactor plate: unequal factors in a row go replica by replica” |
| P3 | errors: every state of a categorical input impossible, every outcome impossible, a linear term in a flat direction, unequal flatness, a mislabeled categorical on a later replica (LP-032) | 27 checks pass: both paths throw | — | — |
| P4 | flat directions in tied clusters over a plate: messages, beliefs, energies against per replica and the closed form; a tilt in the flat direction refused | 27 checks pass after fixes to the probe (an unbalanced parenthesis; (mix = …) without a comma, read as a keyword; a message asked for an internal variable, which both paths rightly refuse) | adopted | ”Flat directions: tied clusters over a plate, batched” |
Found by the suites, not by the review: Lenticulum’s ChainRulesCore extension still marked its old
_as_clusters non-differentiable; with _as_clusters now Mycelium’s (and marked there) the method
was defined twice and overwritten, with a warning on every load. The mark in Lenticulum’s
extension was removed.
Leaning / resolution. Nothing to fix in the change. The review did not run the suites (the maintainer’s session ran all seven packages) nor the benchmarks.