open-problem

Status: closed (2026-10-25). Related: L-SPEC-001 The Factor Interface, L-ADR-0006 Separate likelihood_message from invert, LP-003 SquaredNorm does not check its metric, LP-008 No conformance test for the factor interface, LP-009 Game-level free energy is declared but not implemented.

Problem. A read-only review of master by Codex (2026-10-25, of the 16 commits after 5a850764: the issue #8 energy work, the cherry-picked review-argon fixes, L-ADR-0006) found defects that the test suites passed over, each with a reproduction. Besides reopening LP-003, LP-008 and LP-009:

  1. entropy returned a bare number for the five Lenticulum factors, though its contract makes it an element of the energy space; applying the factor’s own GradedScalarisation to it threw a MethodError.
  2. The nonlinear factor’s likelihood_message and invert disagreed for iterations > 1: only invert relinearised at the posterior, so the BP message was not the site of the posterior (against L-ADR-0006). For , noise 0.1, prior , , four iterations: posterior mean 1.9938 from invert, 2.4634 from the prior times the message.
  3. Predictive LikelihoodFactor free energy set fit and the conditional entropy both to 0, because they cancel in the plain sum; any other scalarisation got the wrong value (ordinal, thresholds , : both are 0.8448).
  4. The mixture likelihood_message (VariationalDiffusion) kept the posterior’s weights after dividing out the anchor; the weights also need each component’s log-partition change (two modes, anchor : weights 0.68/0.32 against 0.59/0.41).
  5. The prior’s cotangent in implicit_factor_pullback was the first unobserved block’s cotangent even when an incoming point overrode the prior: 0.20 where finite differences give 0.
  6. RQ-002 and Universal Approximation for Implicit Diffusion Learners §3 claimed an error from uniform continuity alone; uncentred shifts give .

Resolution. Codex implemented the fixes test-first on codex/review-fixes, one commit each; Claude reviewed them, kept predictive mode working for the count likelihoods (Codex had made it throw) by summing their conditional entropy as a series, and corrected the notes: 1 → 4a49ca5c; 2 → c03014c3, with the graph message on the same routine in bb3af9e9; 3 → c5f07b5a, 3be38b3e; 4 → 4c184538; 5 → 84b69146; 6 → the notes now say , for , and only for centred shifts and . The review is Lenticulum.jl-Review/10_Review_of_master_(Codex).md (outside the repository).

Lesson. Records were marked closed on the strength of commit messages. A resolution names the code or test that closes it, and is checked against the code before the status changes.