A list of data is not a tricolon

[peopleworks]

Closes #100.

rhet.rule-of-three matched any three word-characters, so it could not tell a rhetorical tricolon — "fast, simple, and powerful" — from an author listing three values. #100 found it on "2324, 2425 and 2526" in a support-ticket reply, and measured it: 18 of the rule's 236 corpus matches are lists like that (gene names, doses, year codes), and six texts had no other hit. Reproduced exactly before changing anything.

The narrowing

In the pack, not in C#, so the rule stays editable by anyone. A list is skipped when none of its three items contains a lowercase letter:

\b(?!(?-i:[\p{Lu}\d]+), (?-i:[\p{Lu}\d]+),? and (?-i:[\p{Lu}\d]+)[.,;])(\w+), (\w+),? and (\w+)[.,;]

Two traps shaped it, and both are pinned by tests:

• PatternAnalyzer compiles every pack pattern with IgnoreCase, so an uppercase class would match prose too and silently switch the rule off. The token class is wrapped in (?-i:…). If that ever stops working, "fast, simple, and powerful" stops firing and the test says so.
• \p{Lu}, not [A-Z] — an ASCII class misses ÁREA, ÉPOCA y ÍNDICE. Same lesson as the Turkish dotless ı in The character scanner reads Turkish dotless i as an artifact #62.

Mixed lists still fire (CD31, CD45, and cells): requiring all three to be data is the conservative reading #100 argued for.

The Spanish twin rhet.regla-de-tres gets the same narrowing. The Spanish corpus cannot judge it — 14 matches, none of them data — but the reasoning is language-independent, and leaving the packs asymmetric is how a sibling goes stale.

Published numbers: unchanged in every field

The calibration run was compared field by field before and after the regex change. A baseline run first confirmed the tool reproduces the committed file exactly except engine and measuredOn; the run with the new regex differs from that baseline in zero fields. So published-calibration.json is not regenerated here, and no committee is needed.

humanRatePer1000 stays at 1.09 on purpose. It was measured on a superset of these matches, so it is now slightly generous — and refitting one rule's rate in isolation is the trap #75 describes. It will be refreshed with the next full thresholds run.

A sibling

SKILL.md told agents to "vary the count" on tricolons. In edit mode that rewrites someone's data. It now says a list of data is not a tricolon.

Verification

• Mutation: with the original regexes, all 12 data-list tests fail; with the new ones, all 18 pass.
• 393 + 72 + 7, plus the desktop job's 31.

Thanks to whichever terminal filed #100 — it arrived measured, with a proposal and the open questions named, which is what made this a half-day fix.

🤖 Generated with Claude Code

https://claude.ai/code/session_015PEbbiYSNPw7jE3LrPNhyF