Measurably more correct with Ohmatic in the loop
PCBBench: 62 circuit-design prompts (MIT-licensed, not ours). One generation session per prompt, two checkpoints: the model's first attempt is graded as-is, then the same session continues through Ohmatic's verify→fix loop until it passes or gives up (cap: 10 rounds). Both checkpoints are judged by stock KiCad ERC, an independent judge we don't control, plus a well-formed-document check, so an empty schematic never counts as a win.
Run full-opus48-v3 · model claude-opus-4-8 · judge KiCad 10.0.3
(kicad-cli sch erc) · 2026-07-07. ERC-clean means electrically consistent,
it does not mean the circuit is fit for purpose. Every number traces to the run's
stats.json and
report.md; no hand-typed figures.
Don't take our word for it. One script re-runs KiCad's ERC on every schematic behind the
table above, with your locally installed KiCad. No account, no API key;
nothing runs except kicad-cli on the published .kicad_sch files.
curl -O https://ohmatic.dev/verify.py python verify.py # downloads the run's schematics (~1 MB), re-runs KiCad ERC on each with your kicad-cli, # recomputes the table, and prints REPRODUCED, or the exact divergence. # needs Python 3.10+ and KiCad 9/10.