Section S of the Feature Backlog: edits, moves, gates, scores and a search driver. 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.
| id | feature | status | needs | priority | source | why |
|---|---|---|---|---|---|---|
| S1 | Template edits of plated graphs, stable ids | planned | — | low | L-SPEC-011 Plate Edits | edits and as_plate take about 15% of a search today |
| S2 | Moves and warm starts as library code (beam search, final refit) | proposed | — | low | lib/Lenticulum.jl/examples/structure_clusters.md | the example’s beam with refit reaches the best fit in 62% of the gradient search’s time; a reusable driver would serve other families |
| S3 | Collapsed M-steps as a library feature (integrate out the intermediates of a switch-over-sum) | proposed | — | medium | lib/Lenticulum.jl/examples/structure_clusters.md | the per-factor M-step crawls with nearly deterministic switches |
| S4 | Gated inclusion of edges (spike-and-slab): a binary gate and a slab per candidate edge, inferred with the rest; a structure edit becomes the rounding of a gate | proposed | — | medium | The Regulariser and the Loss §2, §6; Daniel, 2026-11-05 (“valuable but rarely used, or to regularise?“) | separates whether an edge exists from how large it is, so a rare but large edge is not shrunk away |
| S5 | Per-edge diagnostics: usage (posterior mass through the edge), gain when used, net value against its complexity, bootstrap stability | proposed | — | medium | Daniel, 2026-11-05 | tells a rare but valuable connection from one to prune; also says what an out-of-core store must keep loadable |
| S6 | Structure scores beyond BIC: a Laplace evidence (the Hessian at the fit) and a held-out free energy | proposed | — | medium | The Regulariser and the Loss §6 | BIC is the crude Laplace term; held-out scores also judge parts that are not normalised densities |
| S7 | A structure-search driver in the library: move sets (edits, gates, cluster merges), the scores above, beams, restarts and a final refit | proposed | S2, S6 | medium | The Regulariser and the Loss §5–§6; lib/Lenticulum.jl/examples/structure_clusters.md | every search today is example code; merges the “moves and warm starts” row above |