Status: resolved (2026-11-16). The review of P13 (the exactness of a tree sweep checked at run time) ran four probes. Its one finding, that
ExposedFactorandOverlayFactordo not declareexact_messagesand so make a report say “not exact”, is right as an observation and intended: the declaration is opt-in, and delegating to the wrapped factor would claim exactness for an exposed factor whose parameter channels carry uncertain beliefs. No code change. Related: Feature Backlog (P13), passing “Exactness checked at run time”.
Sources: a review of the uncommitted P13 change by Mistral (
vibe --agent review,--max-price 3, $0.98 at list price), in the helper worktreeLenticulum.jl-mistralat86da7a51with the diff applied; the brief, the diff, the answer file and the probes with their outputs, and the repaired probeprobe2_fixed.jl, in../Lenticulum.jl-Review-2026-11-16/p13/. code:lib/Mycelium.jl/src/passing.jl(_inexact_reasons,exact_messages,ConvergenceReport),lib/Lenticulum.jl/src/Lenticulum.jl(the declarations); tests:lib/Lenticulum.jl/test/exact_report.jl.
Problem. The requirement given to the reviewer: report.exact true only when the marginals are
exact, false whenever a step approximated (a mixture collapsed, a MixtureReduction node, an
undeclared factor); the reason names each; iterated schedules never exact; the four-argument
constructor still works; declared graphs without mixtures exact, and their marginals then equal a
closed form. Asked especially for false claims of exactness.
| # | finding or probe | right? | what was done | test |
|---|---|---|---|---|
| F1 | ExposedFactor and OverlayFactor return exact_messages == false although they wrap exact factors; “add methods that delegate to the contained factors” | the observation is right (probe4.out); it is the intended cautious direction, documented in Lenticulum.jl and in passing. The proposed fix is wrong for ExposedFactor: its parameters are channels, and with a non-point belief on them the factor is bilinear in its inputs, its messages not exact; whether they are points is known only at run time. OverlayFactor was not examined | none | — |
| P1 | the test file and its helpers | ran and passed (probe1.out) | — | test/exact_report.jl |
| P2 | a PriorFactor emitting a mixture into a LinearConstraintFactor | the probe errored on its own construction (mixture components declared improper, and an earlier version a disconnected graph). Repaired (probe2_fixed.jl): the report says not exact, “node :obs receives a mixture on :y and collapses it” | — | probe2_fixed.jl only |
| P3 | an ExposedFactor in a tree | did not run past a constructor error of the probe; its conclusion is F1, by reading | — | — |
No false claim of exactness was found. Not probed: a cluster node receiving a mixture, plates
(PlateSchedule is not a TreeSchedule, so it never claims exactness), stale messages from an
earlier run.
Reviewer’s tooling. Its six refused commands all began with cd … &&; its julia commands,
written without output redirection, ran. On P21 the refused julia command had a redirection
(> … 2>&1), which P23 ran; the cause is not settled.
Leaning / resolution. No change. A declaration for wrappers, if wanted, has to look at the run: an exposed factor is exact when every parameter channel holds a point.