open-problem

Status: resolved (2026-11-15). No wrong result, crash or missing check. Mistral reported one finding, that cluster nodes’ block inputs with different categorical labels per replica would be batched where the per-replica path refuses them; its own probe did not run, and repaired it shows the inputs refused. A categorical block holds one set of labels for all its columns, so such inputs arrive as a vector block, which the block path sends through the columns. Not reproduced; the case became a regression test. Related: Feature Backlog (I20), LP-034 Findings of the review of I19, LP-032 Findings of the review of I6.

Sources: a review of the uncommitted I20 change by Mistral (vibe --agent review, --max-price 1; it stopped at that price after 71 turns, with an answer file kept as it went), run with Julia in the helper worktree Lenticulum.jl-mistral at 2d554500 with the diff applied; the brief, the diff, the output, the answer file and the probes (with the repaired ones, -fixed) in ../Lenticulum.jl-Review-2026-11-15/; the logs of the maintainer’s runs of the probes in ~/.cache/lenticulum/i20/mprobe_*.log. code: lib/Mycelium.jl/src/cluster_batch.jl (_block_batchable, _uniform_shape, the block-input factor_messages, _discrete_block, _continuous_block), lib/Mycelium.jl/src/cluster.jl (_continuous_message), lib/Mycelium.jl/src/plates.jl (_group_plan, _group_trees), lib/Mycelium.jl/src/free_energy.jl (_energy_row); tests: lib/Lenticulum.jl/test/regime_plate.jl.

Problem. I20 grouped untied factors without parameters in plate sweeps and free-energy rows, and let cluster nodes read a plate sweep’s inputs as blocks and return their messages as blocks (plates “Untied factors without parameters”, cluster “Block inputs”). The requirement given to the reviewer: the grouped and block paths give the edge-by-edge and column paths’ results exactly, for every input shape, and fall back when they cannot.

#finding or proberight?what was donetest
F1_uniform_shape(::CategoricalBlock, …) compares the columns’ log-probabilities but not their labels, so inputs whose labels differ between replicas (LP-032) would be batchedno: a CategoricalBlock has one labels field for all columns, and belief_block stacks categorical beliefs with different labels into a VectorBlock, for which _uniform_shape returns false. Mistral’s probe probe_label_defect2.jl failed on a missing using LinearAlgebra; repaired, it prints _block_batchable: falsenone in the code”a mislabeled categorical input on a later replica is refused”: the stacked inputs are a vector block, not block-batchable, and factor_messages on them throws
P1the rewritten _continuous_message (a linear scan) against the old one (a Dict), on random inputs: one assignment, several, equal columns, different precisionspassed (probe_continuous_message.jl, rerun by the maintainer: “All tests passed”)—covered by “block inputs give the column path’s messages exactly”
P2block against column inputs for a RegimeFactor with pointspassed after repairing the probe (a reshape of a 2 × 2 matrix to 2 × 2 × 2)—covered by the same testset

Not checked by the review (its answer says so): the I20 testsets themselves, untied state that is non-empty, and fit_em’s key selection. The maintainer’s session ran all seven suites, the two I20 testsets (four mutations of the block path fail 18 to 69 of their tests) and the benchmarks.

Leaning / resolution. Nothing to fix in the change. The price limit, not the turn limit, ended this run: --max-price 1 counts list prices even for the free model (3.9 M input tokens, 97% cached, $1.01), so it stops a review at about 70 turns.