Lab · Experiments
LLM-Assisted Schematic Review
Running
Caught four of twelve known faults. Invented two that were not there.
Tooling / ProcessSample12 bad schematics
Caught4
False positives2
StartedJul 2026
Progress40%
The question
Can a language model usefully pre-screen a schematic before a human reviews it?
Where it stands
Fed twelve schematics with known, documented faults. It found four — all of them obvious ones a checklist would also catch. It missed every subtle fault, and confidently reported two problems that did not exist. As a pre-screen it currently adds review time rather than saving it.
03Bench log
14 Jul 2026
Twelve known-bad schematics assembled from our own history.
29 Jul 2026
First pass: 4/12 caught, 2 false positives.
05 Sep 2026
Netlist input instead of images — no meaningful improvement.
04Next step
Next
Test whether it does better at BOM cross-checking, which is a lookup problem rather than a reasoning one.