Routability reader
Mendicant
Live in this page
Tero’s dynamics
One mould a net
Each tube thickens with the flow it carries and withers without it. A channel two nets want costs both more.
The 22-part board from the Quorra page, with every net but ground growing its own mould. Mendicant runs the same model on a coarser 32 × 32 grid, in Rust.
The foldA board, read as a molecule.
- 01
Backbone
Line the board’s pins up in order around its centre. That sequence is the strand.
- 02
Base pairs
Join each net’s pins by the shortest tree. Each link becomes a chord across the circle of pins.
- 03
Pseudoknots
Two chords that interleave cannot share a layer without crossing. In RNA that pair is called a pseudoknot.
- 04
Displaced
Nussinov’s 1978 RNA-folding algorithm, run with the pairs fixed, counts the nets that must leave the top layer.
On one board, more crossings means more vias from the router, even with wirelength held. That holds on 174 of 250 boards (median rank correlation +0.19, Sep 2026).
On a chipWhere a crossing is a device.
crossings on a 16 × 16 Clements mesh, with all 287 connections routed. Each crossing on a photonic chip loses about half a decibel of light, so the fold counts exactly what matters.
Judged by LiDAR, another group’s open-source photonic router, on its own benchmarks, Sep 2026. Apollo’s code is not public, so its rows are its published table. On the 32 × 32 MMI tensor core, the placement routes all 447 nets with 111 crossings; the layout a person drew has 121.
CompiledAn eighth of a second a placement.
Escape windows
0.123 sa layout, reading every pin’s way out at pad resolution. The Python version took 2.09 s.
17× sooner, Sep 2026The coarse mould
0.026 sa board on a 32 × 32 grid, one mould a net. The Python version took 0.185 s.
7.2× sooner, Sep 2026The same loop
7×10−12the largest relative difference from the Python version, over 131 placements. Every whole-number reading matches exactly.
Held there by tests/test_mould.pyMeasuredWhat it reads, and what it does not.
Board difficultyPins with no way out predict a failed route, on 1,774 held-out boards nothing was tuned on.ReadsAgainst a rulerA count of pads closer than a track and two clearances reads nearly as well. Between boards, most of that headline is pad pitch.ConcededPast the rulerHolding board size and three such counts, the coarse mould still tracks failure.ReadsOne board’s placementsOrdering a board’s clean placements above its failed ones, on 2,045 boards that hold both. Real, and too weak to steer a placer.Too weak6,822 annealed placements of 1,774 open-hardware boards, each routed by two routers under one KiCad check. Sep 2026.
In the build
Mendicant reads boards; it does not route them. It can also grow a power rail. Ask for one, and the tubes that survive become the copper, each sized to the current its loads draw. Otherwise the rail is the hull of its pads.
power_shape: mouldin a design, orSOLDERMASK_POWER_SHAPE- Without the Rust core, the Python version: the same numbers, slower
Written up
The fold and the mould are a paper, reviewed twice. Every reading is scored between boards and within one board. A claim the second review broke is withdrawn where it was made.
- Two routers and one KiCad check decide each label
- The label’s own ceiling measured by nudging placements 0.05 mm
- Read the research notes
Read theboard first.