How we measure the router
On 105 boards a person already routed by hand.
A router can look good on the boards it was tuned on. So we judge ours on boards a person already routed by hand: the 105 real KiCad boards of PCBWorld-Bench's D3 set.
The routing bench routes each of them with a frozen, read-only copy of Provenboard’s engine, and records which commit and which changes it ran.
Three scores per board
- Routed and clean. Is everything connected, and does the copper pass the board's own design rules?
- Next to the paper. The same boards routed by Freerouting, a reinforcement-learning router and an LLM agent, as the benchmark's paper reports them.
- Next to the person. Detours, vias, hard corners, short jogs, pairs and pour islands, ours against the layout a person made of the same board.
The third score exists because a board can be routed and clean and still look like no experienced person routed it.
The rules it keeps
- It never writes to the engine or to a project. Real projects are compared on snapshots.
- The craft score is information, not a gate. A regression is a board that routes worse: an open connection, a failed clean pass, a run that fails.
- Times are indicative. Runs share the machine with whatever else is running.
Every engine change that touches routing runs on the bench before it is committed, and the bench says board by board what changed.