Status: resolved (2026-11-12). One defect, reproduced by a test that failed before the fix and passes after it: a batched plate of cluster nodes read a categorical input with reordered labels on any replica but the first, which the per-replica path refuses. It predates I6 (the shape that decides batching compared the log-probabilities, not the labels). Three other probes, and a second review by GLM-5.3 with five probes, found the batched path equal to the per-replica one. Related: Feature Backlog (I6), L-SPEC-009 Cluster Nodes, LP-031 Findings of the review of the query report.
Sources: a review of the uncommitted I6 change by Mistral (
vibe --agent review), run with Julia in the helper worktreeLenticulum.jl-mistral(2026-11-12). It stopped at its turn limit three times without writing an answer (../Lenticulum.jl-Review-2026-11-12/I6-from-mistral*.md); its probe scripts,scratch/i6/p1_multi_discrete.jl,p3_labels.jl,p4_fallback.jlandp5_graph.jlin that worktree, were run by the maintainer’s session and are summarised here (logs in~/.cache/lenticulum/i6/probe_*.log). code:lib/Mycelium.jl/src/cluster_batch.jl(_input_shape); test:lib/Lenticulum.jl/test/regime_plate.jl, “a mislabeled categorical input on a later replica is refused”.
Problem. Tied cluster nodes over a plate are batched when their inputs have the same shape; the batched methods must then give what the per-replica methods give, refusals included. I6 extended the batching to factors whose coefficients and noise depend on their outcome.
| # | finding | right? | what was done | test |
|---|---|---|---|---|
| F1 | _input_shape recorded a categorical input as (:cat, logp). Replica 1’s joint is built by _cluster_joint, which refuses labels other than 1:K (_check_positional); the other replicas reuse it, so a replica > 1 with labels (2, 1) was batched and read positionally, while factor_message and factor_belief on that replica throw | right: reproduced (p3_labels.jl; the regression test failed 3 of 3 before the fix) | the labels are part of the shape, so such replicas fall back to the per-replica path, which refuses them | the testset named above |
| P1 | a plate of clusters holding a regime factor, a switching factor with a fixed form and a point channel, and a table over two discrete variables (several assignments per outcome group): messages, beliefs, energies, Zygote and ForwardDiff gradients, EM statistics | no defect: batched equals per-replica to | — | p1_multi_discrete.jl (scratch) |
| P4 | inputs of different shape (a point against a Gaussian, different categorical inputs, different observed switch) are not batched and fall back exactly; identical shapes are batched and equal | no defect | — | p4_fallback.jl (scratch) |
| P5 | P1’s model as a graph, plated against unplated: free energy, gradient, EM statistics batched and not | no defect: differences | — | p5_graph.jl (scratch) |
A second review, by GLM-5.3 (vibe --model glm-5-3, same worktree, after the F1 fix, without
access to Mistral’s probes), found no defect. It wrote its answer probe by probe
(../Lenticulum.jl-Review-2026-11-12/I6-from-glm.md) and stopped at its turn limit before its
summary; its last probe was run by the maintainer’s session. Its scripts are in scratch/i6-glm/.
| # | probe | result |
|---|---|---|
| G1 | a cluster row with two varying factors on different switches (one with learned noise, one with fixed), Gaussian priors outside with per-replica means: messages to three channels, beliefs, energies, Zygote and ForwardDiff gradients, moment_statistics batched and not, em_step | equal to |
| G2 | flat directions under per-outcome forms, an observed discrete state equal across replicas, a categorical input with a entry | equal; batched in all three |
| G2b | flat directions where the point matters (, of rank 1, four replicas): energies, messages, pooled statistics | equal |
| G3 | (a varying form with one outcome), positional labels on every replica, two varying factors on the same switch (classes keyed by ) | equal (probe3_corners.jl, run by this session) |
Leaning / resolution. F1 fixed with its regression test. The probes P1, P4 and P5 were not copied
into the test suite; P1’s mix of factors is a case regime_plate.jl does not cover. The review did
not report on Zygote through belief_free_energies beyond P1, nor on device arrays (none are used
by the cluster code).