Stop counting rules that measure the genre instead of the machine

[peopleworks]

Closes the measurement half of #31. The rules that misfire on human academic writing are no longer counted as evidence — and, deliberately, are not hidden either.

What the corpus said

Ninety texts published before generative models existed produced 888 flagged tells: a median of seven per human academic paper, with only two of ninety coming back clean. rhet.rule-of-three fired on half of them.

The scores were never the problem — those rules weigh one or two, the corpus median was 6.8/100 and nothing reached the recommended threshold. The evidence panel was the problem, and it costs this project more than it would cost anyone else. Showing the evidence instead of a percentage is the entire argument, and a teacher who gets seven confident-looking tells on a colleague's paper learns to disbelieve the eighth.

What changed

Each rule now carries the rate at which people use it: hits per thousand words at the ninetieth percentile of the corpus, per language, and only for rules with at least eight texts of evidence behind them. Below that rate a finding is marked AtHumanRate — still shown, still offered to the rewriter, worth nothing to the score. Rules the corpus never sees keep counting on a single occurrence, which is why delve and tapestry are untouched. The mechanism is self-limiting to the genre rules by construction.

Marking rather than deleting is the second attempt. Deleting was written first and it was wrong three ways: evidence vanished as a document grew (padding an essay would bury a tell), the live rewriter lost words it knows how to replace, and it answered "our evidence panel is noisy" by hiding evidence. All three are worse than the disease.

The numbers, reported the way the project reports numbers

Thresholds are fitted on the same corpus this project measures itself against, so the improvement is quoted leave-one-out — each text judged against rates derived from the other eighty-nine:

	before	after
findings per text	9.9	7.0 (−30%)
texts with nothing flagged	2 / 90	9 / 90
corpus median score	6.8	6.1
recommended threshold	25/100	25/100

In-sample it is −32% and 10 clean texts. The smaller number is the one on the page.

rhet.rule-of-three has left the misfire table entirely; the top entry is now rhet.in-terms-of at 10%.

Reproducible, which was the point

dotnet run --project tools/SignsOfAI.Calibration -- thresholds

derives every rate from the corpus and writes it back into the packs. Re-running it produces no diff. The first version of this change pasted numbers from a throwaway script while the code claimed the tool had produced them — an unverifiable "measured" figure, which is precisely what this project criticises in everyone else.

What it does not fix

Rates are per rule, and language models are tuned away from repeating any single tell. The recognisable shape of machine prose is fifteen different tells at one occurrence each, every one of them individually at a human rate. Marking rather than deleting keeps that shape visible and countable, but nothing here scores it. Measuring breadth instead of density needs its own evidence, not a constant chosen today.

And the limit that cannot be argued away: this removes false findings measurably, and cannot be shown to preserve true ones, because the project keeps no corpus of machine-written text — deliberately. What holds it up there is an argument, not a measurement, and it is written where it happens rather than in a commit message.

Also

• HumanRatePer1000 is nullable on purpose: absent means never measured, present-and-zero means a catalog turned it off. A contributed pack that rewords a rule inherits the measured rate instead of silently losing it.
• The CLI distinguishes the two in both output modes (3 signals + 2 at a human rate).
• 9 tests for the gate; 225 core tests, 304 across the solution, all green.