What did the check actually prove?

I asked my agents to spend half my weekly allowance in one hour. The logs counted 284 million tokens, and 96.66% of them were cache reads. Across all my retained logs, $33,807 of use at list price, and about a third of it could not be tied to a product.

I put my Garmin heart rate and sleep over every coding agent session I have on record: 168 million tokens. With agents writing my average heart rate was 64.6, without them 66.2. Sleep after the heaviest days was 11 minutes shorter, which is noise. One person, 81 days, and three numbers I got wrong first.

Jev said yes to 62 of 63 events. I also tried it on contradictory notes, corrections and my own decisions. Seven graphs, some useful answers, and a much shorter list of jobs I would give it.

Almost everything I write now is a prompt, typed to an agent, and none of it is meant for a person. So I graded six weeks of it, which is all the deletion timer left me. Prompts where I wrote down what done meant left something on disk 65% of the time against 39%. Then I checked whether knowing that had made me better, and it had made me worse. Then I found 193 days of my own record had already been deleted by a default I had never opened.

One night of coding agents left ten surfaces carrying a sentence their own cited source contradicts, across nine repositories, and not one was caught by the agent that wrote it. In the clearest case a merchant tool announced that no published figure priced an obstacle; the figure was row four of the same table it was already quoting two rows from. Then the piece itself said nine instances over a list of ten, and only a reader who had not written it counted the bullets.

The practice without the story: one pass whose only job is to open every cited source, and which is not allowed to do anything else. Four steps, the four shapes that keep recurring, the one-line grep, and the three things that do not work: tests, rendering it and looking, and being careful.

Prompts are the least edited words I produce, and they steer everything. So I graded six weeks of mine. The ones that said what done looks like left something durable behind 65% of the time, 64 of 99. The ones with no stated intent, 39%, 392 of 1,001. On a rerun six days later, counting only real commits instead, it is 30% against 17%. The same command runs on your logs.

I gave six agent skills away and spent a week worrying I had sold my edge. Wrong worry. My whole working context lives in markdown I wrote, every model I ask reads those same files, and then they all agree with me.

I put six of my agent skills in a public repo under an MIT licence and handed them to strangers. Then I stopped and asked the obvious thing: what did I think was mine in the first place.

Twenty done-claims, each re-derived from the artifact it was a claim about. Eighteen grounded, zero contradicted, two unverifiable. The two it refused to score are the reason the zero means anything.

Tomorrow I was going to file a provisional patent. This morning I checked where the specification's own reduction to practice lives, and found I had shipped it to a public repo four days earlier. The open-source instinct and the patent instinct were both mine, in the same week, in the same directory, and they never met.

The guard was right. The number it validated was true, of an object the real system had never seen. A live position ran at 5x cross-margin while the audit log honestly reported leverage 1. Nothing failed, nothing lit red.