Negative Result Taxonomy

A failed attempt is a symptom to read, not a place to stop.

The team that produced this rule had built dozens of candidates and measured a handful of them. Deaths were terminal by default: a gate said no, the line stopped, and the next session started something new. Across roughly fifteen deaths, one diagnosed rebuild was on the record — and it produced the best artifact they owned, improving four separate metrics at once from a single measured cause and a single precedented fix, for about an afternoon's work.

The loop worked. Almost nothing required it.

There is a second lesson sitting on top of the first. The founding document's headline count of what had been built and measured turned out to be wrong in both directions — a later census, the deliberate read across every attempt, found more work had reached measurement than the headline claimed and more loops had already run than anyone had credited, because nobody had written them down as loops. The census is the mechanism, and it corrected its own charter within a day of that charter being written.

That is the shape of the problem: taken singly, every abandonment looked careful — specification first, thresholds written down, an honest verdict — and nothing about it is visible from inside any one attempt. It only appears when you count across attempts, which is why the count has to be mandatory rather than occasional.

Three classes, one test

The question is what actually failed — the whole approach, this attempt at it, or the apparatus: