A checker for KiCad boards. It compares your design against the manufacturer's datasheets, and it tells you what it didn't check.
I'm looking for three boards to run it on.
Tatsu Ikeda, October 2026. Seven slides. Most of what's on them is pasted straight out of the tool or the repo.
My test board. 13 parts, 50 by 40 mm, with defects planted on purpose. The next slide is what the tool says about it.
$ hwci check tests/fixtures/demo rules run: 19 (waived findings: 0, no waiver configured) COVERAGE parts considered: 2 (reference prefixes the pin-map rule skips are not counted) pins compared against a datasheet pin table: 0 NOT CHECKED: 2 (U1, U2) NOT CHECKED for errata: 2 (U1, U2): no errata record was read for these parts, which is not the same as the vendor publishing none NOT CHECKED, no rule exists: layout: no rule in this pack judges placement or routing quality NOT CHECKED, no rule exists: signal integrity: no rule in this pack computes impedance, crosstalk, or timing NOT CHECKED, no rule exists: EMC: no rule in this pack evaluates emissions or susceptibility [cut here: the sha256 of every input file, and a false positive count for every rule] error decoupling-proximity U2 U2 pad 5 on power net '+3V3': nearest capacitor C2 is 6.38 mm away (limit 3 mm) error decoupling-proximity U1 U1 pad 2 is on power net 'VDDA' which has no capacitor at all error floating-pins U2 U2 pin 3 (EN) is unconnected: no wire, label, or no-connect marker error i2c-pullups U1 I2C line 'SCL2' has no pullup resistor to a power net (bus cannot idle high) warning pin-map-conformance U1 U1 (HWCI-MCU10) has no datasheet-backed pin map on file, so its pin assignments are unverified [6 more findings cut] 7 error(s), 2 warning(s), 2 info not checked: 4 finding(s) (2 warning, 2 info) report what this pack could NOT verify, not a defect it found
Run this morning on commit 284aae6. Lines cut and columns narrowed to fit; nothing reworded.
The top block is the part I care about. Before it lists what's wrong, it lists what it couldn't look at. On this board that's both chips, because neither has a datasheet on file.
ERC and DRC compare your design to itself. If the symbol is wrong, everything downstream agrees with the wrong symbol. So each part gets a file with the pin table copied out of the vendor's PDF, plus which PDF it was.
[provenance] document = "SBOS392I" revision = "I" revised = "2026-04" url = "https://www.ti.com/lit/ds/sbos392i/sbos392i.pdf" sha256 = "39839abc9bf09417d0a4b446c6e922679f6dc63b19072e284169dd23007e65fe" retrieved = "2026-07-29" page = 4 table = "Table 5-1. Pin Functions"
parts/ref3312.pinmap.toml. When a symbol disagrees with the table, the finding names the pin and the document.
I have 17 of these. That's the whole list of parts it can verify today. Five of them say at the top that no person read them:
# PROMOTED 2026-09-01. TWO INDEPENDENT AGENT READS OF THE # RENDERED PAGES, UNANIMOUS WITH THE MAP BELOW, AT THE # OWNER'S DELEGATION. # NOT INDEPENDENTLY HUMAN-SIGNED. [...] # That is retrieval, not recall, twice over, which is the # defect the rule exists to stop. It is not a human's eyes, # which is the other thing the rule buys. A human re-check # remains worth doing and has NOT happened.
parts/mcp6002.pinmap.toml, header. "Agent" means an AI model reading page images.
43.6%
That's how much it catches.
I wrote down 47 ways a board design can be wrong. Then I put each one into 25 real open-source boards and ran the checker. It caught 262 of 601.
The idea isn't mine. It comes from an engineer who did this for circuit boards at Bell Labs starting in 1977. He emailed it to me in September.
Fault model 1.0.1, corpus run of 5 Sep 2026. Some of the misses sit behind KiCad's own checks, which this test doesn't run yet.
| Fault, as written in the model | Caught |
|---|---|
| A copper zone on a net is drawn but never filled | 25 of 25 |
| A track width, via drill, annular ring or zone clearance is below the published floor of the fabricator | 25 of 25 |
| A regulator's input or output rail has no capacitor | 22 of 22 |
| A control pin (EN, RESET, MODE, BOOT) is connected to nothing | 20 of 20 |
| An I2C bus has no pull-up resistor to any power net | 19 of 19 |
| A connector's pinout is wrong or inverted | 0 of 25 |
| A differential pair is wired with its polarity swapped | 0 of 24 |
| The part-number field holds a distributor catalogue code | 0 of 24 |
| A support component is present but of the wrong type | 0 of 22 |
| A capacitor's voltage rating is at or too near the rail it sits on | 0 of 21 |
Ten of the 47. Sixteen have never been caught once. The full list is in the repo and you can have it.
The tool keeps a hash of every datasheet PDF it cites. On 31 August I had it download six of them again and compare.
| Part | Document | Hash on file | Hash today | Revision |
|---|---|---|---|---|
| TPS61023 | SLVSF14B | b181cbe9e2d3 | 1750450d5e3d | B, unchanged |
| TLV9062 | SBOS839N | c37698e10c1c | 57cfc69310fe | N, unchanged |
docs/calibration/2026-08-31-hash-recompute-gap.md. Four other documents came back identical.
TI changed both files and kept the same revision letter. If you track datasheets by revision, which is what I'd have done, you see nothing.
The changelog inside the new TLV9062 file says what they did:
Added TLV9062S RUG (VSSOP) package, pinout information to Pin Functions: TLV9062S table.
The pins for the package I use didn't change. For the TLV9062 that check was a text comparison, not me rereading the pages.
I wrote the code with an AI coding assistant, and AI readers did some of the datasheet work on slide 3. The checker itself has no model in it. Same files in, same report out.
I scraped GitHub for KiCad projects to test on. There are only a few thousand and a lot of them are unfinished. 4,049 run. It's like pushing rocks up a hill. I didn't think it would be this difficult.
Send me one KiCad project you're about to fab. NDA if you want one.
I run it and send you the report within a week. It'll look like slide 2, for your board. If your parts aren't in my 17, it will say NOT CHECKED and I'll tell you which ones I can add.
Free. Or a fixed fee if you'd rather pay for it, your call.
Then tell me if it was useful and what it got wrong.
Tatsu Ikeda hello@testcoupon.dev
If you know a shop that should see this, forward it. testcoupon.dev