open-problem

Status: resolved (2026-10-26), option A. RatioFactor emits a likelihood potential; posterior reweighting remains explicit. Related: LP-006 Gaussian entropies are relative to Lebesgue measure, LP-008 No conformance test for the factor interface, LP-009 Game-level free energy is declared but not implemented, L-SPEC-001 The Factor Interface, ratio.

Sources: original repository analysis; code: lib/LenticulumCore.jl/src/open_model.jl, lib/Mycelium.jl/src/messages.jl, lib/AdversarialFactors.jl/src/ratio.jl, lib/AdversarialFactors.jl/test/conformance.jl, lib/Mycelium.jl/test/runtests.jl

Theory (CT-ML wiki): Bayesian Inversion · Statistical Game

Defect and closed form

The former outgoing particle message reweighted the cavity and was therefore a posterior. Pooling it with the cavity failed (two SampleBeliefs have no product rule); a supplied target input could override the cavity. The graph energy also used the incoming weights instead of the factor-belief weights.

For and ,

On equally weighted particles , the posterior weights are . The incoming energy expectation is zero, whereas the factor-belief expectation of is .

Resolution

  • AbstractPotentialBelief is the extensible improper likelihood interface; LogPotentialBelief(logφ) wraps a callable. Log-density is defined up to an additive constant. There is no sampling, mean, or entropy contract. Richer subtypes can expose gradients, parametric structure, or batched evaluation.
  • RatioFactor’s graph and lens likelihood messages return this potential without reading the cavity or target inputs. invert and explicit reweight return the posterior. The callable captures parameters/state; density evaluations discard returned network states. Use inference-mode state and do not mutate captured values while messages are used.
  • Mycelium reweights particles through belief_logdensity, adds log-potentials, preserves Dirac dominance and Trivial unit laws, and refuses arbitrary potential products with Gaussians, categoricals, or mixtures. It does not resample. Callable identity gives zero distance; other evaluators give Inf. Damping potentials is an inspectable no-op.
  • Mycelium’s free-energy call passes incoming variable→factor messages . The unary factor belief is . RatioFactor now averages under that reweighted belief. This is the Bethe energy contribution only: no continuous particle entropy is supplied (LP-006), so it is not a complete continuous free energy. Dirac game quantities remain; particle game methods still throw for lack of a chosen reference measure.

Verification

lib/AdversarialFactors.jl/test/conformance.jl replaces the three broken LP-016 checks with passing closed-form assertions for pooling once, target exclusion, and posterior energy expectation. It checks the lens likelihood separately from inversion and tests a particle prior with two RatioFactors on one variable in both insertion orders. The marginal weights are proportional to the prior times once.

lib/Mycelium.jl/test/runtests.jl checks the potential products, unit and Dirac laws, unsupported parametric products, finite-support associativity, distance, damping, and method ambiguities. Ambiguity detection exposed the pre-existing (nothing,nothing) distance ambiguity; an explicit Inf method now resolves it.

Full package suites run with julia --project=<package> -e 'using Pkg; Pkg.test()':

packagepassfailerrorbroken
LenticulumCore236000
Mycelium494000
Lenticulum (top level)775000
AdversarialFactors254000
ImplicitLayers316000
AlgebraicFactors14396000

These counts include the LP-017 changes already present in the working tree. The Mycelium suite includes Test.detect_ambiguities(Mycelium, LenticulumCore; recursive=true); no Aqua suite is configured in these package tests.