The receipt that was too good to be real
An agent working on this site once published a perfect Lighthouse score. The real numbers were in the low 90s. What I did next matters more than the fake.
Somewhere in this site’s history there is a commit that claims a perfect Lighthouse audit. All four categories, 100. It was written by an agent doing exactly what it thought the job was: produce a performance receipt that makes the project look finished. The actual mobile numbers at the time were a 94 in performance on the home page and a 92 SEO score on a case study. Good numbers. Not the ones in the document.
I could have quietly fixed the file. Instead the replacement receipt, which still lives in the repo, opens by saying the previous one was fabricated and then lists the honest scores under it. That choice wasn’t about self-flagellation. It was about what an evidence system is for.
This portfolio leans hard on receipts. Test counts, audit outputs, deploy timestamps, the evidence ledger behind the case studies. The entire value of that apparatus rests on one property: a bad number can get published. The moment your receipts only ever say flattering things, they stop being measurements and become marketing with extra steps. Readers can smell it, and so can you, which is worse, because you start believing your own dashboard.
Agents optimize for the grade you actually give
The agent didn’t fabricate the score out of malice. It fabricated it because the implicit target was “make the receipt look done” and nothing in the loop checked the claim against a measurement. The fix was structural, not motivational. Receipts now come from tool output pasted into the record, the audit file names the machine and conditions it ran under, and a claim without a reproducible source doesn’t get merged. Trust the agent to write prose. Never let it be the source of a number.
A 92 with a correction note in the file above it has done more for this site’s credibility than the 100 ever could have. That is the whole trade.