Problems found while specifying or building Lenticulum.jl, numbered
LP-nnn(Moonkale usesP-nnn, SophiaSP-nnn). Each problem gets a note made from the problem template (vault/templates/Problem Template.md) and a line here. Closed problems stay listed, with what closed them.
Implementation notes also record what their code does not do; a problem gets an entry here when it needs a decision, crosses several files, or is waiting for someone to pick it up. GitHub issues are linked from the problem they belong to.
| # | Problem | Status |
|---|---|---|
| LP-001 The vault build uses GNU-only sed | sed -i in docs/site/build.sh fails on macOS (issue #9) | closed (2026-10-23; CI guard 2026-10-25) |
| LP-002 The Jacobian side is not implemented | DI factor/graph derivatives; DEQ dense/Krylov IFT and dense HVPs; fixed-state SciML; realised flooding/residual/damped replay and opt-in CPU message fixed-point adjoints | batch C (2026-10-28), seven suites / 25,857 assertions and both docs builds pass; measurements in LP-002; native replay/device implicit/general divergence/metric limits remain; Core Enzyme NaNs traced to upstream forward mode (Enzyme.jl#3884) |
| LP-003 SquaredNorm does not check its metric | An indefinite M makes the Jensen gap negative, silently | closed (2026-10-25, reopened once) |
| LP-004 link_sources rewrites unrelated notes | The bibliography linker drops links in three unrelated notes | closed (review-argon, 2026-10-23) |
| LP-005 Learning the noise covariance | Opt-in full/diagonal precision parameters for Q and Σ; normalised free energy with Wishart/log-normal noise priors or an eigenvalue floor | resolved (2026-10-26; ForwardDiff/Zygote CPU paths) |
| LP-006 Gaussian entropies are relative to Lebesgue measure | Categorical KL entropy of a continuous belief is infinite; ours are relative to Lebesgue | accepted, documented |
| LP-007 Missing implementation notes | gaussian.md and beliefs.md are cited but not written | closed (review-argon, 2026-10-23) |
| LP-008 No conformance test for the factor interface | Common checker covers every declared exact domain, independent product games and evaluated diffusion bounds; all 22 production types and eight test fixtures audited with deliberate semantic boundaries | resolved 2026-10-27 |
| LP-009 Game-level free energy is declared but not implemented | Labelled diffusion ELBO bounds; exact affine/source, scalar Plucker, degree-one polynomial and Gaussian composition/product games; other domains throw with reasons | resolved within agreed scope 2026-10-27 |
| LP-010 Where the graph of a problem comes from | Pipeline note and Gaussian single-factor search vs graphical lasso; measured BIC/Bethe limitations; chain recovery test | resolved within note + example scope (2026-10-27) |
| LP-011 Gradient couplings and isfrozen are declared but unused | Per-edge cuts, subset freezing, realised schedule gradients, CPU implicit graph adjoints, terminal score estimates and straight-through factor | extended in batch C (2026-10-28); CPU/JLArrays replay subset, native backend/implicit device/estimator limits remain |
| LP-012 Factor graphs and schemes | Cited/proved theory note: RLCT parameter evidence, dual-number tangents, linear sheaf consistency | resolved within theory-note scope (2026-10-27) |
| LP-013 Defects found in the review of master | Codex’s review of master: entropy type, iterated nonlinear site, predictive fit, mixture weights, prior cotangent, a rate claim | closed (2026-10-25) |
| LP-014 Emitting edges drop the cavity message from the free energy | Cavity-reading factors now reject Emitting() edges during graph free-energy evaluation (uses_cavity, option 2); marginals unchanged | closed (2026-10-26) |
| LP-015 Momentum had the wrong sign on its decay | Momentum and Nesterov used instead of (Cruttwell et al. Example 3.16) | closed (2026-10-25) |
| LP-016 RatioFactor emits a particle posterior as a likelihood | Extrinsic log-potential messages, particle pooling once, target exclusion and posterior energy contribution checked | resolved (2026-10-26) |
| LP-017 DiffusionFactor omits supported polarities | Complete capped support, lazy traversal and joint unobserved-block inversion | resolved (2026-10-26) |
| LP-018 Defects found in the review of the older packages | Codex’s review of VariationalDiffusion, top-level, AlgebraicFactors and AdversarialFactors: Poisson entropy hang, RED-Diff calibration factor, per-column σ in energy-network Jacobians, batched VJP times, hard clamps, MixtureProx maxima, covariance validation, ordinal cancellation, weight overflow; four wrong note claims | resolved (2026-10-29), reviewed by Claude |
| LP-019 Defects found in the review of the core packages | Codex’s review of Mycelium, LenticulumCore and ImplicitLayers: relay free energy, mixture operand order, one-triangle Gaussian validation, finite-difference steps, damped tree convergence, overlap cancellation, pullback overflow, scalar particles, zero residual budget, norm underflow; two wrong note claims; an LP-018 symmetry check too strict | resolved (2026-10-29) by Claude; not yet reviewed |
| LP-020 The fixed-message parameter gradient is not the free-energy gradient | free_energy_gradient with diagonal couplings held messages, not beliefs, fixed: not the free-energy gradient (−0.146 against 0.679 on a two-variable tree), and alternating parameter_step stopped away from maximum likelihood (a = 0.754 against 1.179). Fixed by holding each factor’s belief fixed (factor_belief, belief_free_energy), the envelope gradient | resolved (2026-11-01) |
| LP-021 The coupled gradient drivers cut diagonal edges at fixed messages | The coupled drivers (adjoint = :implicit, non-diagonal couplings) still cut a diagonal edge at its fixed message, so on a tree with default couplings adjoint = :implicit is off by 0.58 from the converged free energy’s gradient, while the default driver (fixed beliefs, LP-020) and ExactCoupling everywhere agree | resolved (2026-11-02): fully cut factors at fixed beliefs in the coupled drivers (option 1); mixed factors keep fixed messages |
| LP-022 Constraint factors are normalised in their residual, not their channels | A constraint factor is a density of its residual, not of its channels, so with every coefficient learnable the free energy on data prefers A = 0 (F = −57.01 against 32.34 for least squares on 200 points) and, with learnable noise, has no lower bound. output = :y fixes the data channel’s coefficient and makes it a normalised conditional density | mitigated (2026-11-04): output option; a warning in free_energy_gradient |
| LP-023 Findings of the external reviews of 2026-11-04 | Two source-level reviews (ChatGPT, Gemini): eight confirmed defects (the likelihood_message default, skipped EP updates stored by flooding and selected by the residual schedule, Dirac–categorical contradictions, categorical labels read by position, empty sample beliefs, per-message parameter indexing, singular implicit pullbacks, the Laplace covariance of non-conservative fields) | partly resolved (2026-11-04): the eight fixed with tests; proposals recorded |
| LP-024 Findings of the review of 2026-11-09 | A review that ran the suites (Claude): partly flat Gaussian messages threw or gave wrong marginals on a tree (a smoother with position-only measurements), freeze paths froze the wrong coordinates, composition of Gaussian factors failed for non-diagonal noise, damping dropped the improper flag, parameter_step and free_energy_gradient disagreed; four corrections to the theory notes | partly resolved (2026-11-10): fifteen fixes with tests on refactor; scaling and element-type items open |
| LP-025 Findings of the review of I5 | Mistral’s review of the I5 change (trees read by position), with Julia: parameter_step refused an optimiser state given as a Dict; a suspected per-key scan of the energy rows not reproduced | resolved (2026-11-12): the Dict accepted again, with a test |
| LP-026 Findings of the review of I18 | Mistral’s review of I18 (cheaper per-message checks), with Julia: the loop symmetry test returned false for a non-square matrix instead of throwing; _mean_cov’s error named belief_cov; no test of the memo’s weak references | resolved (2026-11-12): both fixed with tests, the tests added; a race needing a data race on A not reproduced |
| LP-027 Findings of the review of the I18 plans | Mistral’s review of the planned factor → variable messages (the rest of I18), with Julia: same results, errors and warnings as the old path on random graphs and 16 threads; a claimed compile-time regression; no concurrency test | resolved (2026-11-12): no defect; the compile-time claim not reproduced (the old path compiled per input set too); a concurrency test added |
| LP-028 Findings of the review of P1 | Mistral’s review of P1 (where the flagged and the extended Gaussian forms meet), with Julia: a ForwardDiff crossing left open, a test that did not build what it said, an overclaiming header | resolved (2026-11-12): the defect fixed; the pre-existing limit is @test_broken (P21) |
| LP-029 Findings of the review of P17 and P18 | Mistral’s review of the moves of mixture reduction and sample moments, with Julia: identical results before and after; stale notes, two internal functions of one name | resolved (2026-11-12): no code defect; notes fixed, one function renamed |
| LP-030 Findings of the review of round_trip | Mistral’s review of round_trip (F3), with Julia: a MethodError for a precision vector of element type Any; untested edge cases | resolved (2026-11-12): fixed with a regression test; soft evidence, views and Float32 now tested |
| LP-031 Findings of the review of the query report | Mistral’s review of query_report (P11), with Julia: the wrong ε for :ill_conditioned; :not_lowest by rank, not energy (missed at rank 0, raised on a tie); an Inf round trip unflagged without a tolerance | resolved (2026-11-12): four fixes, each with a regression test |
| LP-032 Findings of the review of I6 | Mistral’s probes of batched per-outcome cluster nodes over a plate (I6), with Julia: a categorical input with reordered labels on a later replica was read positionally (predates I6); three probes equal | resolved (2026-11-12): labels in the batching shape, regression test |
| LP-033 Findings of the review of the ridge residual | Mistral’s probes of ridge_residual and the :off_ridge flag (F3), with Julia: no wrong result; four missing tests (symmetrisation, positivity of normal curvatures only, the Inf flag, the default codimension) | resolved (2026-11-13): each with a test that fails on its mutant |
| LP-034 Findings of the review of I19 | Mistral’s probes of the stacked-joint batching of tied cluster nodes over a plate (I19), with Julia: batched against per-replica messages, beliefs, energies and errors on random instances, flat directions, unequal rows; no defect (the review wrote no answer; its probes, fixed, are the evidence) | resolved (2026-11-13): two probes adopted as regression tests |
| LP-035 Findings of the review of I20 | Mistral’s review of untied data-factor groups and block inputs to cluster nodes (I20), with Julia: one finding (mislabeled categorical blocks batched), not reproduced since a categorical block has one set of labels; two probes equal | resolved (2026-11-15): nothing to fix, the case a regression test |
| LP-036 Findings of the review of I21 | Mistral’s review of the mixture block with a precision per component and the kept energy rows (I21), with Julia: point × mixture blocks of different sizes not refused | resolved (2026-11-15): refused with DimensionMismatch, test |
| LP-037 Findings of the review of P23 | Mistral’s review of Zygote through the constraint and switching cavity free energies (P23), with Julia: one inv(_chol(...)) left in likelihood.jl | resolved (2026-11-16): replaced; claimed impact not reproduced |
| LP-038 Findings of the review of P21 | Mistral’s review of ForwardDiff through the flat path with only η dual (P21): no probe ran (commands refused); its probes’ failures were their own errors; switching and two-input gaps now tested | resolved (2026-11-16): no defect; a dropped partial of η along null(Λ) documented |
| LP-039 Findings of the review of P13 | Mistral’s review of the run-time exactness check of a tree sweep (P13), with Julia: wrappers (ExposedFactor, OverlayFactor) not declared exact | resolved (2026-11-16): intended (opt-in; delegation would be wrong for exposed parameters); no false exactness claim found |
| LP-040 Findings of the review of P24 | Mistral’s review of ForwardDiff and Zygote through a cluster node’s free energy at fixed messages (P24), with Julia | resolved (2026-11-16): no defect; two probes weaker than reported, covered by the suites |
| LP-041 Findings of the review of P26 | Mistral’s review of ForwardDiff and Zygote through flat directions of a cluster’s joint precision (P26), with Julia | resolved (2026-11-16): no defect; one probe without assertions, covered by the suites |
| LP-042 Findings of the review of P25 | Mistral’s review of ForwardDiff through the flat path when its flat directions move (P25), with Julia: nested duals gave a MethodError | resolved (2026-11-16): fixed, regression test |
| LP-043 Findings of the review of P4 | Mistral’s review of ForwardDiff and Zygote through extended Gaussians and the replay through a hard relation (P4), with Julia: Zygote threw on a belief with no Gaussian direction | resolved (2026-11-16): fixed, regression test |