Rules that fire on human academic writing: rule-of-three at 50%

[peopleworks]

The calibration in #30 produced a list nobody had before, and this is what to do about it.

Every rule below fired on text published before generative models existed. Each hit is a false
positive by construction — there is no judgement call to make.

Rule	Texts it fired on (of 90)	Share	Total hits
rhet.rule-of-three	45	50%	100
stat.burstiness	25	27.8%	25
lex.moreover	23	25.6%	67
lex.furthermore	23	25.6%	55
rhet.in-order-to	22	24.4%	60
lex.just	19	21.1%	30
lex.facilitate	17	18.9%	29
rhet.in-terms-of	16	17.8%	30
lex.comprehensive	14	15.6%	36
lex.ademas	12	13.3%	28

The problem

A rule firing on half of human academic prose is measuring the genre, not the machine. Scholars
were writing "furthermore" and grouping things in threes long before anyone trained a transformer.

What this is not

Not an argument for deleting them all. Some of these genuinely are AI tells and genuinely common in
human academic writing, which is exactly why they carry weight rather than certainty, and why the
catalog explains each one. The score is a reading of prose and prose is arguable — that was always the
deal.

But rhet.rule-of-three at 50% is not carrying information. It is background.

What would close this

For each rule near the top, one of:

1. Reweight. Lower its contribution so it colours the score instead of driving it.
2. Narrow it. rhet.rule-of-three presumably fires on any three-item list; the AI tell is closer
   to abstract triples in a summarising sentence, not "we measured height, weight and age".
3. Demote to Info. Shown in the report, contributing nothing to the score — the catalog entry
   still teaches the writer something.
4. Retire it, and say so in the catalog.

The discipline this needs

Re-run the calibration before and after and put both numbers in the PR. Changing rules to make our
own number look better is exactly the failure mode this measurement exists to prevent, so the change
has to be argued on the rule's merits and the number reported either way.

Note also that lowering weights lowers the whole score distribution, which moves the recommended
threshold down with it. That is fine — the threshold is derived, not chosen — but it means "the
false-positive rate improved" is not by itself evidence of anything.

Related

• Measure how often this flags a human, and publish it #30 for the measurement and the method
• Baseline: calibrate the outside-the-range count against known authorship #29 for the calibration still owed to the per-writer baseline
• Spanish academic writing remains the most wanted corpus contribution; lex.ademas is the only
  Spanish rule in the top ten and the Spanish side currently rests on encyclopedia prose alone.

[peopleworks]

Fixed in 383032a ("Stop counting rules that measure the genre instead of the machine") — verified in main with git merge-base --is-ancestor. The current calibration page (engine 0.4.0) reflects it: rhet.rule-of-three now fires on 6.7% of the human corpus, not 50%; the worst remaining offender is stat.burstiness at 27.8%, which the page publishes. This issue stayed open a week after the work shipped — closing against the code, not the memory of it.