Status: resolved (2026-10-26). Complete capped enumeration, predicate-based structural checks and joint inversion on unobserved blocks. Related: LP-008 No conformance test for the factor interface, L-SPEC-001 The Factor Interface, factor.
Sources: original repository analysis; code:
lib/LenticulumCore.jl/src/channels.jl,lib/VariationalDiffusion.jl/src/factor.jl,lib/VariationalDiffusion.jl/src/implicit_factor.jl,lib/VariationalDiffusion.jl/test/conformance.jlTheory (CT-ML wiki): Bayesian Inversion
Problem. The predicate accepted every assignment with an unobserved channel but the list omitted latent and multiple-target splits. Inversion returned only the first unobserved block. Three broken membership tests and one broken dimension test recorded it.
Resolution. Option A: supported_polarities now returns all default-precision
Observed/Unobserved/Latent assignments through arity four (65 entries). One-target/all-other-
observed entries retain their order at the front; the remainder follows lazy traversal.
Above four channels it throws ArgumentError pointing to supports_polarity and
polarities. The predicate remains the authority. The lazy iterator is type-unstable,
since each assignment has its own type, and complete traversal still costs exponential work.
Core hasnopolarity and isunidirectional retain list-based defaults; diffusion overrides
both without enumeration. Mycelium’s issink delegates to hasnopolarity. Graph polarity resolution constructs each per-target assignment from available messages
and checks it with the predicate; it does not need a list.
invert and likelihood_message return a joint belief on concatenated unobserved blocks
in channel order. Point messages select the prox solution; Gaussian and mixture messages
restrict the existing full-state approximation. Likelihood messages divide out the
product of diagonal target anchors. An improper Gaussian is allowed; an improper mixture
component throws because its normalized weight is undefined. The diffusion prior remains
in the message. factor_message(f, target, …) remains per-target. Joint prior points use
the same concatenation; shared-block points retain their prior convention.
Checks. The four broken assertions now pass in “Diffusion factor partial conformance”,
with no LP-017 polarity_gap or message_gap. For the isotropic fixture the field
integrates to : the site has precision , so multiplying by
gives . Both target coordinates and the covariance are checked.
“Diffusion polarity enumeration and cap (LP-017)” checks counts at arities two, three and
four, equality with lazy traversal, ordering and structural answers above the cap.
“Joint diffusion messages (LP-017)” checks noncontiguous unequal blocks, joint-prior
cotangents, point/Gaussian/mixture messages, anchor pooling and correlated covariance.
“Joint mixture rejects undefined anchor division (LP-017)” checks that removing the only
confining anchor throws instead of assigning normalized weights to flat components.
“Joint mixture anchor pooling (LP-017)” checks recovery of both components and weights
when a two-target mixture site is multiplied by its diagonal anchor. Core “Lazy polarity
enumeration (LP-017)” checks lazy traversal independently of a list.
Game gaps remain in LP-009; this does not claim arbitrary-prior or stochastic-backend
conformance.
Validation run (2026-10-26)
Full suites, using julia --project=lib/<Package>.jl -e 'using Pkg; Pkg.test()'
(and --project=. for Lenticulum):
| Package | Pass | Fail | Error | Broken |
|---|---|---|---|---|
| LenticulumCore | 230 | 0 | 0 | 0 |
| Mycelium | 459 | 0 | 0 | 0 |
| Lenticulum | 775 | 0 | 0 | 0 |
| VariationalDiffusion | 1143 | 0 | 0 | 0 |
| ImplicitLayers | 316 | 0 | 0 | 0 |
| AdversarialFactors | 232 | 0 | 0 | 3 |
| AlgebraicFactors | 14396 | 0 | 0 | 0 |
VariationalDiffusion’s original testset passed 889 assertions in 6m50.1s; its conformance file passed 254, including the joint and cap testsets above. The remaining three broken assertions belong to Ratio particle conformance (LP-016).