A technician photographs every nameplate on an install and this page checks those photos the way a quality desk would before the job closes out. AI reads each plate, the photo's own text confirms each value, and code checks every value against the job's design. A person sees only what comes back uncertain or failing and settles each one in the queue.
| Who | What it does | How it's checked |
|---|---|---|
| AI | Reading each plate's fields from the photo | Scored field by field against the answer key |
| Code | Finding the words on the photo and their boxes on the owner's Mac | Its words are stored with each photo |
| Code | Confirming each read by the photo's own text near its label | Found, partly found, or not found |
| Code | Comparing each value with the design and the other photos | Unit tests with parity tests against the page's copy of the rules |
| Code | Sending reads to a person from the confidence line | Scored as errors caught against reads sent |
| A person | Settling what's uncertain or failing | The queue and the audit log |
Every plate is made up, and no real maker, serial, or place appears on any photo. Scripts already in the code would rebuild a client's real photos from their pixels with no metadata. A person would confirm every mask and key value before any model saw them.
The scan runs in the build and again on the public copy of the repo. It checked 116 published image files and found 0 problems. It looks for any metadata, any photo missing from the masking log, and any word the text recognition can still read under a mask.
A read is found when its exact value is in the photo's own words by its label or carries its own unit where a unit says enough. A value by another field's label never counts and a minimum circuit ampacity read as the largest breaker goes to a person. A model number one character off is only partly found. Case, spacing, and units are evened out before the comparison. An O never turns into a 0.
These are the installer's own design rules. They're kept as data and run by the same code in the build as in this page. They aren't an inspection or an electrician's sign-off on the work. Edit a design value beside any check on the job and it reruns on the spot.
Each model got the same photos and saw none of the test photos until the prompt was frozen. No AI graded its own reading.
| Model | Fields right | Wrong reads past the line | Wrong verdicts that skipped the person | Reads sent to a person | Cost per photo | Seconds per photo |
|---|---|---|---|---|---|---|
| Claude Sonnet 5.5 | 99% | 0 | 0 | 1% | $0.0146 | 6.1 |
| Claude Haiku 5.5 | 99% | 0 | 0 | 1% | $0.0008 | 4.7 |
| GPT-6 Luna | 99% | 0 | 0 | 1% | $0.0006 | 6.8 |
The page opens on GPT-6 Luna because it's the cheapest model that cleared every bar on the tuning photos. The rule was set before any model read a photo and it never looks at the test photos because a lucky test score shouldn't be what picks the model.
| Field | Fields right | Values in the key |
|---|---|---|
| Model | 97% | 29 |
| Volts | 100% | 29 |
| Phase | 100% | 16 |
| Amps | 100% | 17 |
| Min. circuit | 100% | 8 |
| Max breaker | 100% | 10 |
| Capacity | 100% | 16 |
| Refrigerant | 100% | 8 |
| Breaker | 100% | 6 |
| Shot | Fields right | Values in the key |
|---|---|---|
| shot at an angle | 100% | 38 |
| shot in glare | 95% | 20 |
| shot in low light | 100% | 20 |
| shot straight on | 100% | 61 |
A wrong read is dangerous only when the photo seemed to confirm it and the table counts those at each position of the line.
| Line at | Sent to a person | Wrong reads caught | Wrong reads past the line |
|---|---|---|---|
| Only what's not found | 1 (1%) | 0 | 0 |
| Partly found and not found | 2 (1%) | 0 | 0 |
| Every read | 139 (100%) | 0 | 0 |
Each read was recorded once on the owner's machine with its prompt, tokens, cost, and time. This page only replays them.
| Photo | Model | Recorded | Tokens in and out | Cost | Seconds |
|---|---|---|---|---|---|
| p01 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 1038 | $0.0174 | 7.0 |
| p04 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 856 | $0.0156 | 6.1 |
| p06 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 379 | $0.0108 | 3.3 |
| p08 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 840 | $0.0154 | 7.3 |
| p11 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 781 | $0.0148 | 6.7 |
| p13 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 366 | $0.0107 | 2.9 |
| p15 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 338 | $0.0104 | 2.7 |
| p17 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 368 | $0.0107 | 3.5 |
| p19 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 987 | $0.0169 | 8.3 |
| p21 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 976 | $0.0168 | 8.6 |
| p38 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 581 | $0.0128 | 5.3 |
| p40 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 467 | $0.0117 | 4.7 |
| p42 | Claude Sonnet 5.5 | 2026-11-09 | 3505 and 592 | $0.0129 | 5.3 |
| p01 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 1199 | $0.0009 | 6.7 |
| p04 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 911 | $0.0008 | 3.9 |
| p06 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 807 | $0.0008 | 3.6 |
| p08 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 1078 | $0.0009 | 5.3 |
| p11 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 748 | $0.0007 | 3.6 |
| p13 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 844 | $0.0008 | 4.1 |
| p15 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 839 | $0.0008 | 3.9 |
| p17 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 900 | $0.0008 | 4.1 |
| p19 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 1373 | $0.0010 | 5.9 |
| p21 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 1014 | $0.0009 | 4.8 |
| p38 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 679 | $0.0007 | 3.2 |
| p40 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 526 | $0.0006 | 2.5 |
| p42 | Claude Haiku 5.5 | 2026-11-09 | 3505 and 734 | $0.0007 | 4.0 |
| p01 | GPT-6 Luna | 2026-11-09 | 3006 and 878 | $0.0007 | 10.0 |
| p04 | GPT-6 Luna | 2026-11-09 | 3006 and 610 | $0.0006 | 5.7 |
| p06 | GPT-6 Luna | 2026-11-09 | 3006 and 580 | $0.0006 | 5.8 |
| p08 | GPT-6 Luna | 2026-11-09 | 3006 and 882 | $0.0007 | 9.0 |
| p11 | GPT-6 Luna | 2026-11-09 | 3006 and 557 | $0.0006 | 5.6 |
| p13 | GPT-6 Luna | 2026-11-09 | 3006 and 532 | $0.0006 | 5.7 |
| p15 | GPT-6 Luna | 2026-11-09 | 3006 and 495 | $0.0005 | 5.3 |
| p17 | GPT-6 Luna | 2026-11-09 | 3006 and 536 | $0.0006 | 5.7 |
| p19 | GPT-6 Luna | 2026-11-09 | 3006 and 636 | $0.0006 | 6.3 |
| p21 | GPT-6 Luna | 2026-11-09 | 3006 and 624 | $0.0006 | 6.3 |
| p38 | GPT-6 Luna | 2026-11-09 | 3006 and 429 | $0.0005 | 5.1 |
| p40 | GPT-6 Luna | 2026-11-09 | 3006 and 309 | $0.0005 | 3.5 |
| p42 | GPT-6 Luna | 2026-11-09 | 3006 and 400 | $0.0005 | 4.5 |
The pattern carries to any plate, form, or label your technicians photograph. It takes a few hundred of your photos with their answers confirmed by your quality desk, your design rules written as data, a run of the models on photos they never saw, and a line set where your team is comfortable. After that your people see only what the photos can't settle and the rest of the work moves on its own.
It reads what the photo shows and a plate behind a bracket or under glare goes to a person. It checks the installer's rules and nothing else. It isn't an inspection. It isn't production software. The page replays reads recorded in advance and can't read a new photo.
Five files carry most of the idea and the whole project is below them.