tighten-style

tighten-style

Cuts what does not earn its place and gives paragraphs a spine, against the catalog in references/style-rules.md. Every finding cites a rule ID, so the report reads as a copyedit keyed to a document you can argue with rather than a list of opinions.

The rules are ours. They cover the ground Strunk covered and the ground he could not — he wrote before markdown and before a machine could produce fluent prose with nothing behind it.

The floor

Tightening stops at the target register's own density. Shorter is not the target. Cutting past the measured function-word ratio manufactures the clipped, aphoristic register that reads as machine-written — the failure filter-tells calls overshoot. Where a writing-voice/ corpus exists, match-structure supplies that floor; without one, stop at each rule's stated threshold.

The floor is venue-keyed (GH-338). By default it is the author's own density. With --venue <name>, the venue profile (writing-voice/venues/<name>.yaml, see the writing-voice rule) supplies it instead: the profile's targets (sentence_length_mean/stdev) become the sentence floor — a whitepaper tightened to newsletter density reads clipped — and its hedge_policy keys the TS-08 threshold: zero flags every hedge (book voice, rule 10), minimal flags pairs, calibrated keeps single calibrated hedges on empirical claims and only flags stacks. An explicit --sent-floor still wins over the profile. No --venue, no change.

Never touch a direct quotation, a normative requirement, a citation, or a number.

Procedure

python3 <tighten-style>/scripts/check_style.py <file> [--json] [--hedge-policy zero|minimal|calibrated]
python3 <tighten-style>/scripts/tighten.py --article <file> [--venue <name>]   # the rewrite, via Ollama
python3 <tighten-style>/scripts/tighten.py --article <file> --check-only   # plan without model calls

1. Run the checker for the findings and the reading pass for the judgment rules (TS-06, TS-07, TS-09, TS-11) it cannot see.
2. Run the tightener for the rewriting. It selects the instead→do pairs for the rules each paragraph fired, prompts the second model family with the pairs — never the rule prose — and gates every candidate with match-voice's verify step. Paragraphs that fail the gate keep their original text.
3. Read the register markers it prints (passive, agentive, nominalization, connectives, before → after). Rising markers on a shrinking draft mean the pass moved toward the assistant register, which is the failure this design exists to prevent.
4. Do not tighten by hand-applying the rules with the drafting model. The full evidence is in references/hand-tightening-regression.md. GH-222 measured what that does: a paper excerpt moved from distance 26.1 to 6.5 from the AI-draft fingerprint under a faithful rules pass, overshooting the draft's own passive rate. The rules read as instructions by an instruction-tuned model ARE that model's register. The catalog stays for understanding and for the checker; the pairs are its delivery.