It would be cheating to keep the old price

2026-08-21

The agent-readiness audit is now €4,300 and two weeks. It was €6,500 and two to three weeks. Nothing came out of the scope to pay for that.

The scanner is the same one, isitagentready. The manual checks that sit over the scanner are the same. The re-scan after the fixes is still the thing the engagement is judged on, and it is still the only claim I make about the result.

What moved is that the part a client used to pay for twice is now written down once.

What a client used to pay for twice

An audit has two halves. One half is measuring a surface: run the scanner, read what it says, check by hand whether the number is telling the truth. That half is real work and it stays.

The other half was me deciding, again, what a passing row actually looks like. Twenty-two checks, each one an open specification that moved in the last year, and for each one the question was the same: what does a correct implementation of this look like, what does the scanner accept, and what is the concrete change that flips it. I answered those questions from scratch every time, and the client paid for the answering as well as the measuring.

That is the half that is now a file. The checklist has 42 rows. Every row names the pass condition, points at a working implementation of that row on a live surface, and links the fix recipe. There are 25 recipes behind it, one per check or close group.

The live surface the rows point at is this site, because it passes every scored check and I can show the code for each one.

What it cost to get there

I read my own workspace file by file over seven days in August and recorded 367 findings. That was written up at the time. Since then the same treatment went to the four repositories at byte level, and then to the folders and documents, and then three verification rounds ran over the folder audit and the session before it.

The folder and document audit removed 2 317 files and 116 622 446 bytes. Not one line of source was in what went. It was temporary renders, build caches, a dependency tree and six revision copies of one file. I am not quoting a before and after total for the workspace, because three measurements of it taken the same day differ by up to 39 281 bytes: tools write into the folder that was supposed to be the anchor, so that total is not a number anyone should cite.

Three sessions did that work: 2 254 messages, and about 366 million tokens as the raw sum. Weighted the way billing weights them, where a cache read is a tenth of a fresh input token and an output token is five, it comes to about 88 million.

Across every session recorded on this machine the raw sum is 7 197 173 453 tokens, and the weighted figure is 1 849 081 095. Those totals are a floor rather than a count. The transcript tree holds 25 days, and a disk restore in August took five weeks of it with it, so the missing period is missing from the number too.

That spend is mine. It produced no invoice and it was never billable. What it produced is the file that means the next audit does not repeat it. Keeping the old price and the old clock would mean charging a client for work that is already done.

Why the accuracy does not drop

The obvious reading of a lower price and a shorter clock is that something got smaller. What happened is the opposite, and the checklist proved it on me before it proved it on anyone else.

Writing the checklist out found a hole in my own material. The scanner scores a check called ard, the capability manifest at /.well-known/ai-catalog.json. My own recipe index named 21 checks and that one was not among them, and no recipe covered it. My own gate did not catch it either, because the gate reads the checks a recipe claims and has nothing to say about a check with no recipe at all. An audit run from that index would have skipped a scored check and still looked complete.

That is one check out of 22, on a site that serves the manifest correctly and passes the check. It was invisible for as long as the method lived in my head, and it was visible as soon as the method became a list.

A shorter engagement also removes a real source of error. Two to three weeks is long enough for the specifications to move underneath the report, which has already happened here. A payment specification dropped in a point release the extension that three of my guides described, and the MCP server card proposal moved onto an extensions track while the sentence about it stood still. Two weeks leaves that less room. It does not remove it.

What these numbers do not prove

Between the two windows the transcript tree records, the weighted cost per message fell from about 69 700 to about 41 100 tokens. It is tempting to read that as the cleanup paying for itself, and I am not going to claim it. The two windows are five weeks apart, the models in them are different, and the mix of work is different. The drop is real in the record and its cause is not established, so it is an observation and not evidence.

The checklist has also not yet run against a client surface. It was built from 13 rounds of reading my own, which is a real basis and a narrow one.

What is unchanged

Fixed scope, written before any payment. Async only, no calls and no calendar links. Read access is enough, production credentials are not requested, and write access is scoped per task only if implementation is bought separately. The report names 3 to 10 fixes in order of impact, and the next scan either reads higher in the categories the report named or the report explains which tradeoff was kept on purpose.

The other prices did not move. The Shopify agent storefront check is €1,900 within 48 hours of a written kickoff, advisory is €3,000 per month with a three month minimum, and implementation is €1,500 per day. All prices exclude VAT.

The scope is agreed in writing first. I reply by email within one business day.

Related

• I thought it was a small job
• Choosing an agent-readiness audit
• Measurement-led agent readiness