overview

Section P of the Feature Backlog: what the refactor review of 2026-11-10 asked for on the three paths; it comes before every other section. Open rows only; a row that is done moves to Backlog Done 2026-11 with its id. Status, priority, ids and the column needs are explained in Feature Backlog, which also holds the dependency graph.

Sources: original to this repository; the sources of each row are in its source column.

After the refactor (Refactor 2026-11) work on the paths comes before every other section of the Feature Backlog. The rows here are the ones that review asked for; rows of other sections that serve the same end are named. The structure-search driver (path C) is S7.

idfeaturepathstatusneedssourcewhy
P2Hard inputs in the free energies of SwitchingConstraintFactor, the plate paths and the frozen beliefs of the gradientA, Cproposed—gaussian “Free energy with hard inputs”refused or untested there
P3A TrivialBelief() input (no message) to GaussianFactor takes the exact flat path, as an everywhere-flat ExtendedGaussian doesAproposed—gaussian “Extended Gaussian inputs” (2026-11-11)with the shortcut returns no message where the exact one fixes off ; changes the first sweep of every schedule, so measure it on the existing tests first
P5HardConstraintFactor with a singular nonzero noise in the free energy (hard along its null space, soft along its range); learnable (its term needs an SVD, which ForwardDiff lacks)A, Cproposed—hard_constraint “Free energy” (2026-11-11)refused today
P6HardConstraintFactor in plates (a factor_messages block method) and in MeanField() nodesA, Cproposed—hard_constraint “What it does not do”a plate of hard constraints runs per replica; mean-field nodes refuse it
P8Element type in, element type out: the graph free energy in the element type of its messages, with a testApartial: LenticulumCore (2026-11-10)—LP-024Float32 graphs are promoted to Float64 in the total
P9A relative convergence criterion for Gaussian messagesAproposed—LP-024the bound on natural parameters is absolute
P10noisefree_nodes against field_nodes on the calibration benchmark; a default for the nodesBproposedF2Implicit Diffusion Learners §5 (2026-11-10)nodes with noise have twice the smoothing bias at a level
P12density_lambda as the default for non-point messagesBpartial: a warning (2026-11-10)—LP-024uncalibrated covariances reach the graph silently
P14A shared solver layer (root finding, the implicit-function adjoint, one tolerance rule, one report), shared by VariationalDiffusion and ImplicitLayersBproposed—Refactor 2026-11three implementations with three conventions; ImplicitLayers’ tolerances are absolute and written for Float64. Its first fix, now that it is back in lib/
P15A coverage report; a job that runs the examples and benchmarks at small sizeallproposed—review of 2026-11-09which branches no test reaches is unknown; examples and benchmarks are run by nothing (Aqua.jl in every suite: done 2026-11-10)
P16Owned scalarisations: one graded energy, a scalarisation chosen per parameter key (each block descends its own), for games such as a GAN; with a stability check of the joint Jacobian—proposed—GANs as Two Factors §4 (2026-11-10)every learnable node descends the one scalar , so no game is expressible; a per-factor sign is the zero-sum special case
P27A Gaussian mixture times an extended Gaussian in Mycelium’s product (_mixture_product), for a SwitchingConstraintFactor (mixture or EP rule) on a variable a hard relation reachesA, Cproposed—P4 probe (2026-11-16), test/extended_plates.jl (broken)ArgumentError: no product rule for mixture components, in plates and single graphs