Status: resolved (2026-11-15). One finding, right and fixed: the new shortcut “a point block times a mixture block is the point” did not check that both blocks have the same number of replicas, so blocks of different sizes gave a block of one of the sizes. Inside the library a plate’s rows always have one size, so no sweep met it; the shortcut now refuses such blocks with a
DimensionMismatch, and a test shows it. Related: Feature Backlog (I21), LP-035 Findings of the review of I20.
Sources: a review of the uncommitted I21 change by Mistral (
vibe --agent review,--max-price 3, finished after 53 turns, $0.83 at list price), run with Julia in the helper worktreeLenticulum.jl-mistralat4ffd955cwith the diff applied; the brief, the diff, the output, the answer file and the probes in../Lenticulum.jl-Review-2026-11-15/i21/; the maintainer’s run of its probe in~/.cache/lenticulum/i21/mprobe_req1.log. code:lib/Mycelium.jl/src/blocks.jl(ComponentMixtureBlock,combine_blocks,_same_replicas),lib/Mycelium.jl/src/cluster_batch.jl(_continuous_block),lib/Mycelium.jl/src/free_energy.jl(_energy_rows); tests:lib/Lenticulum.jl/test/regime_plate.jl.
Problem. I21 added a mixture block with a precision per component, returned by plates of tied cluster nodes, and kept the plate’s energy rows in the graph per store layout (blocks “Mixtures with a precision per component”, free_energy “Batched energies over plates”). The requirement given to the reviewer: the block’s columns, distance and products equal the per-belief rules, the block path equals the per-replica path, and the kept rows equal rows found afresh for every store.
| # | finding or probe | right? | what was done | test |
|---|---|---|---|---|
| F1 | combine_blocks(::PointBlock, ::Union{MixtureBlock,ComponentMixtureBlock}) and the reverse order return the point block without checking the replica counts; the column path truncates in one order and throws a BoundsError in the other | yes, for blocks of different sizes (none arise in a sweep). A test refusing them failed 8 times on the change as reviewed (test_combine_nreplicas.jl shows the behaviour) | both methods check the counts and throw DimensionMismatch | ”ComponentMixtureBlock: columns, distance and products against the per-belief rules”: blocks of and replicas, both block types, both orders |
| P1 | replica of a ComponentMixtureBlock against hand-built per-replica mixtures, proper and improper | its checks pass; the probe then built a block whose logw had the wrong shape, which the constructor refused (the probe’s error, not the change’s) | — | covered by the I20 block testset (which now returns this block, improper components included) |
| — | block_distance, the energy-row cache, unchanged free energy and EM | ”passed by code analysis”, not run by the reviewer | — | the maintainer’s tests: the distance against belief_distance with only the precisions moved, the cache across stores of one and of another layout, the closed form; four mutations of the change fail 2 to 57 tests |
Not checked by the review (its answer says so): several threads using the cache at once, detached edges, a whole EM run. The maintainer’s session ran all seven suites and the benchmarks.
Found by the maintainer while testing the cache, not by the review: a plate schedule on a store
whose block rows are numbered otherwise than MessageStore(g)’s gives wrong results silently (P22
in the Feature Backlog); it predates I21.
Leaning / resolution. F1 fixed with its test. The other operations that take a point block
(combine_blocks(::PointBlock, ::GaussianBlock)) have the same missing check and the column path
truncates; left as they were.