soldermaskdocs

Research

Two workstreams share this tree. The platform is the pipeline, the command line and the app — everything under The pipeline. The research is this: the corpus of routed boards, the model that reads it, and the instruments built to ask whether any of it helps.

It is written up separately because it answers to a different standard. A platform change is right when the board is better; a research claim is right when it is measured, on how many boards, against which baseline, and with what it does not show stated beside it.

  • Routability — the corpus, the congestion maps, and the model the placer ranks with. It carries a standing correction block at the top naming which of its numbers are withdrawn.
  • Physarum — a slime mould routes the placement in the anneal's time, and what its saturation can and cannot tell you. It carries a correction block: the "ceiling" it once gave for the within-board null is withdrawn.
  • Photonics — a one-second placer for feed-forward photonic chips: no crossings on the interferometer meshes, and fewer crossings and less waveguide than the layouts shipped with the benchmark (8% fewer crossings on the 32×32 tensor core). It places five of LiDAR's eight chips, and its correction block says why this is not "beats Apollo".
  • The rubric — the placer against the people who made the boards, on what a reviewer actually grades.
  • Multi-agent RL — a literature review, parked: nothing published beats a tuned annealer on a real router, and the one cheap test that would settle it here.
  • Routers — Lightcycle against the A* router on the 300-board referee bench, and what each router is told.

The write-ups are paper/isonet.tex and paper/mendicant.tex, and CHANGELOG.md under Unreleased is the running log: every experiment, its numbers, and what it concluded. Several of the obvious experiments have been run and came back null; they are written up as fully as the wins.

Since September

Research logged after these pages were written, newest first. Each line is the bolded claim of a CHANGELOG.md entry under Unreleased, with its date; the entry has the protocol, the counts and what it does not show.

DateThreadWhat it found
16 Nov 2026Datasheet records as contractsNull. Checked at every tolerance corner, the records catch 14 of the 38 September faults the simulator and the design check both miss, against a bar of 15, and alarm on 13 of 122 good boards against a bar of 2.
15-16 Nov 2026Layout gymOn 26 open-source KiCad boards of 4 to 10 layers, re-laid and judged under each board's own rules, the people who made them pass 14. On the authors' placements soldermask passed 1, then 4 after a night's fixes with Lightcycle's exact margins (four passes held on a second run). On a placement of its own it passes one, lg-2ac413b851.
9 Nov 2026Datasheet reader A/BNull. 96 of 106 datasheet clauses passed with the reader off and 97 with it on, over twenty fresh tasks (p = 1.0).
8 Nov 2026Our placer against CypressCypress (NVIDIA and Cornell, ISPD 2025) on 60 held-out boards, six layouts a board, the same router: ours routes 48 and Cypress 31; Cypress routes none ours fails (p = 1.5e-5).
8 Nov 2026Simulator fault injectionOn 144 September designs it had never seen, the simulator caught 96 of the 119 injected faults in the classes it targets (81%), with a false finding on 3; the design check caught 4.
7-8 Nov 2026Design-check A/BsShown the new check, designers met the target requirement on 13 of 15 tasks against 8 of 15 (p = 0.06). On 16 fresh tasks graded by an independent author it did not repeat: 9 of 16 in both arms (p = 1.0).
16 Oct 2026BendNull. A Lightcycle-shaped kernel ran 600 to 1,940 times slower than plain Rust, gained nothing from threads, and twice returned a wrong answer.
13 Oct 2026PCBGolfOur last entry scored 13,533, 4th of 20: five single-layer boards, no vias, every net joined.
5 Oct 2026Four-layer routing benchOn 141 held-out boards their authors routed on four layers, Lightcycle passes 92 (65.2%) and the next open-source router 51 (36.2%), paired 43 to 2.
4 Oct 2026Flow boundNull. On all 10 held-out boards every router left open, a fractional routing fits at every cell size tried, so no certificate of unroutability exists to find.
1-2 Oct 2026Router leversNull. History schedules, per-net noise, net ordering, CMA-ES on the settings, forced topology and an exact path pool each close at most one of the bench's near-miss boards.
27 Sep 2026Referee benchRe-judged on tasks that draw each board's outline and holes, Lightcycle closes 230 of 300 boards and the A* router 138. See Routers.
26 Sep 2026Escape oracleWithdrawn. The author's own escape copper does not double what the router finishes: counted as a build counts it, 8 of 58 fine-pitch boards, the router alone 8.

The standing rule

Between-board and within-board are different questions, and on this corpus they have had different answers.

Between boards is: can this score tell an easy board from a hard one? Within a board is: can it order one board's own anneals? The second is the one the placer actually asks — it has six placements of the same design and has to pick one — and a strong between-board AUC has repeatedly turned out to be board difficulty wearing a new name. A score that cannot rank a single board's anneals cannot help the placer, whatever its AUC.

So every number here is reported both ways, a partial correlation holding board size sits beside any raw one (most raw signals on this corpus are a size prior), and the payoff is given in the units the pipeline cares about — router runs saved, boards routed clean — next to the headline.

Two more habits worth knowing before reading a number:

  • learn status reports training day, not today. soldermask/assets/routability.json carries the metrics computed when the model was fitted and prints them straight back. They do not move when the corpus does. See Benchmarks.
  • Check that a test set is external before calling it external. The last one was not: freerouting vendors PCBench inside its own benchmark fixtures, and 866 byte-identical copies were found in the training directory on 5 Sep 2026. A second audit on 12 Sep found 2,050 of PCBench's boards in the training bank in files KiCad had re-saved, and the external figure of the time was withdrawn. The bank now refuses a board by its content.

On this page