Status: resolved (2026-11-16). The review of P24 (ForwardDiff and Zygote through a cluster node’s free energy at fixed messages) ran six probes and found no defect. Two probes were weaker than reported (below); what they were meant to show is covered by the suites. Related: Feature Backlog (P24, P26), cluster “Derivatives at fixed messages”.
Sources: a review of the uncommitted P24 change by Mistral (
vibe --agent review,--max-price 3, $0.41 at list price), in the helper worktreeLenticulum.jl-mistralatc2284852with the diff applied; the brief, the answer file and the probes in../Lenticulum.jl-Review-2026-11-16/p24/(the probes’ outputs went to its session only, not to files). code:lib/Mycelium.jl/src/cluster.jl,cluster_batch.jl,ext/MyceliumForwardDiffExt.jl,ext/MyceliumChainRulesCoreExt.jl,lib/Lenticulum.jl/src/regime.jl; tests:lib/Lenticulum.jl/test/regime_plate.jl.
Problem. The requirement given to the reviewer: ForwardDiff and Zygote differentiate
local_free_energy of a ClusterFactor and of a RegimeFactor at fixed messages in ps, the
belief’s dependence on ps included, per node, through scalar_free_energy and on a plate with
block rows; Float64 values and the batched paths unchanged; flat directions refused with an
ArgumentError, never a wrong number. Asked especially for dropped or wrong Zygote tangents.
| # | probe | right? | what was done | test |
|---|---|---|---|---|
| P1 | a regime factor with Gaussian messages: ForwardDiff and Zygote against central differences, fixed-belief gradient differs | passed, as the testset does | — | test/regime_plate.jl (P24 testset) |
| P2 | the clustered mixture of factor analysers through scalar_free_energy on block rows | passed | — | test/regime_plate.jl (MFA testset, seed 1) |
| P3 | Float64 values unchanged | the probe checks only that values are finite, not that they equal the previous ones; equality of the batched and per-replica Float64 paths is tested by the existing testsets, which pass | — | test/regime_plate.jl, test/cluster.jl |
| P4 | flat directions refused | passed: ArgumentError under both | — | @test_broken in the P24 testset |
| P5 | a cluster of {prior of x, regime factor} with a Gaussian message on z | passed, as the testset does | — | P24 testset |
| P6 | Float32 parameters | reported passing; not rerun here. With Float32 leaves _numtype promotes to Float64, so the batched path stays on | — | probe6.jl only |
No defect was found. Not probed: tied parameters across clusters of a graph without plates, a
cotangent on one graded part only (the testset weights all three), Zygote with messages that
depend on ps (they get a @not_implemented tangent by design).
Reviewer’s tooling. The brief gave the exact julia command line without cd and without
redirection; every probe ran. It wrote the answer file as asked and finished within its turns.
Leaning / resolution. No change.