Status: resolved (2026-11-12). Two defects were found and fixed with tests, and one gap in the tests was filled. One suspicion was not reproduced, and one finding predates the change. Related: Feature Backlog (I18, P19), LP-025 Findings of the review of I5.
Sources: a review of the uncommitted I18 change by Mistral (
vibe --agent review), run with Julia in the helper worktreeLenticulum.jl-mistral(2026-11-12). Its answer is in../Lenticulum.jl-Review-2026-11-12/i18-mistral.md, and its scratch tests arescratch/i18_core_review.jl,scratch/i18_memo_review.jlandscratch/i18_device_review.jlin that worktree. code:lib/LenticulumCore.jl/src/gaussian_belief.jl,lib/LenticulumCore.jl/ext/LenticulumCoreChainRulesCoreExt.jl,lib/Lenticulum.jl/src/gaussian.jl; notes:lib/LenticulumCore.jl/src/gaussian_belief.md, “Per-message costs”;lib/Lenticulum.jl/src/gaussian.md, “What the checks cost”.
Problem. The change was required to leave every result unchanged. It covered a loop
_nearly_symmetric, a fused _symmetric_part, _mean_cov (the mean and the covariance from one
factorisation), and a memo of _rank_deficient keyed by the array’s identity. Mistral found these
clean, with run output for each:
_symmetric_part, bit for bit (42 000 cases, ForwardDiff duals among them);_mean_covagainst the two functions, for values, Zygote and ForwardDiff;_mean_covon JLArrays withallowscalar(false);- the memo under in-place changes, after it overflows, and with 8 threads (40 000 calls);
- the suites of LenticulumCore, Mycelium and Lenticulum.
| # | finding | right? | what was done | test |
|---|---|---|---|---|
| F1 | _nearly_symmetric of a non-square dense matrix returns false; the generic method throws DimensionMismatch (so Gaussian(zeros(2), zeros(2,3)) threw ArgumentError instead) | right, reproduced | a non-square matrix goes to the generic method | LenticulumCore test/i18.jl, “the loop agrees with the generic test”: both throw DimensionMismatch for 2×3 and 3×2 |
| F2 | _mean_cov’s error names belief_cov, while the old _moments failed in belief_mean | right, reproduced (same type, different message) | the check names belief_mean | test/i18.jl, “_mean_cov from one factorisation”: the message equals belief_mean’s for a declared-improper, a singular and an asymmetric belief |
| F3 | no test that the memo leaves arrays to the collector, or that a reused object id is caught | right (a gap; Mistral’s scratch test found both properties hold) | tests added | Lenticulum test/i18.jl: after GC.gc, some of 200 dropped arrays have dead weak references. Two planted entries under an array’s id, one with another array’s reference and one with a dead reference, both carrying the wrong answer: the memo recomputes. A random search for a reused id found none in 20 000 tries, so the reuse is planted |
| F4 | the SVD runs before copy(A), so a change to A from another thread between the two stores a new copy with the old answer | not reproduced; it needs a data race on A itself, which is already undefined in Julia | nothing changed | — |
| F5 | a wide A with a NaN returns true where svdvals throws | not a change: the old _rank_deficient also returned true for m < n before looking at the values | nothing changed | — |
| F6 | gaussian.md and the backlog not updated | right; they were not written yet when the review ran | updated with the change | — |
Leaning / resolution. Fixed with a test where reproduced. F4 is recorded only.
Not checked: Enzyme through _mean_cov and the memo (no test drives Enzyme through a Gaussian
belief).