Ai vision

Boards to test against

Two of them, and the difference between them is the point.

The end-to-end checks run over both (boards in integration_test.go), because a tool that behaves on one and not the other has been tested on one. The first is what the tool says about unfinished work; the second is the first board on which address/command U4->U5 -- the second fly-by leg -- exists as real copper to match, which is why tuning gains 116.8 mm there against 61.1 mm here.

ai-vision.kicad_pcb is the pinned fixture. The tests do not merely run against it, they encode facts about it that were expensive to establish: that its address bus has 27 nets of which none makes the second fly-by hop, that KiCad reports 687 violations and 499 unconnected items before the tool touches it, that its DQS3 pair is 1.479 mm out. Those numbers are the regression test.

So do not replace it. Add a board alongside it instead.

Running against another board

Everything takes a path:

make report BOARD=demo-pcb/your-board.kicad_pcb
make tune   BOARD=demo-pcb/your-board.kicad_pcb

or upload it in the web front end, which needs no setup at all:

make serve

The .kicad_pro should travel with the board wherever it goes. Without it every clearance falls back to the board-wide minimum rather than the net classes the designer set, and the tool says so rather than quietly working to looser rules. A .kicad_dru beside it is noticed too, and reported as not interpreted.

What a third board would be good for

demo-pcb-2 covered the gap that mattered most: a fly-by chain routed far enough to match per leg. Two things are still untested on real copper.

A board fully wired -- every DDR net joined, and the other interfaces too -- would let the result be judged the way a layout engineer would judge it, with KiCad's unconnected count at zero before and after. Everything the tool reports today is read against a board that is partly holes, and "499 unconnected" is a number the user has to hold in their head while reading every other one.

A board with no DDR at all is also worth having, and is not a special case: the tool reads every interface it recognises on any board, and only the DDR half is skipped. Ethernet, USB and MIPI are detected and measured on both fixtures here, but neither board has one of them far enough along to be lengthened.

Add either alongside, under demo-pcb-3/; the fixture rules in .gitignore already match demo-pcb*/, so only the design files are committed, not KiCad's 101 MB of vendored libraries or its autosave archives.