It took the struck-through number.

Last Tuesday I put a number in print and labelled it a guess. This week I went and measured it — ten rate confirmations, two models, sixty runs, and all the code published so you can check my work.

I was wrong about the cost. I was wrong about which question mattered. Both corrections are below.

— Aman

Last week I estimated half a cent to have a cheap AI model read a rate confirmation, called it arithmetic rather than measurement, and promised to go measure it.

Done. Ten rate confirmations, three passes each, thirty runs, no failures. A third of a cent per document — my estimate was 63% high, in your favour. 97.5% of fields came back correct.

Then I looked at which ones didn't. The rate. The broker's MC number. The load number.

The machine is close to flawless at everything that costs you nothing.

I built ten rate confirmations in ten broker layouts — the bordered fax, the dot-matrix printout, the email forward, the legal one with the rate buried in a paragraph, the two-pickup one, the one where somebody crossed out the printed rate and wrote a new one beside it. Then I degraded them so they look faxed, and ran each three times through Claude’s cheapest model, scoring twelve fields against known answers.

The limit, before the number: these documents are synthetic. I wrote them to mirror real broker layouts. They are not real rate cons out of a carrier’s inbox, and the accuracy figure is not a forecast for your paperwork.

Cost: $0.00337 a document. Accuracy: 97.5% of fields. Nine of 360 reads were wrong — all nine in three fields: the rate, the broker’s MC, the load number.

Every error repeated on all three passes. They are deterministic. Running it again does not fix them.

It took the struck-through $1,475 instead of the $1,650 handwritten beside it, every time. It returned the carrier’s MC where the broker’s belonged. It read TQ as TO.

Then I ran the same corpus through a better model. It got all three right and made no factual errors at all — 360 for 360. Price: $14.95 per thousand documents against $3.37.

Which retires the cost argument. At two hundred rate cons a month, the expensive option is three dollars.

Still owed from Issue #9: the 50-state CDL self-reporting map. Being built. Late.

I expected cost to be the story. It wasn’t. Both models cost approximately nothing; the difference between them was three fields — the ones that become an invoice, a vetting record, and a payment reference.

The typing is cheap. The money is worth reading.

Next Tuesday: Compliance Tech — September 8, ten named devices, and what it means that seventy-nine ELDs have come off the registered list in nineteen months.

— Aman · [email protected]