VoiceDNA

VoiceDNA: 13 elements

VoiceDNA is a local profile of observable writing mechanics. Build it from at least two non-empty samples from one writer. It compares a draft with those samples; it does not identify an author or determine whether AI wrote a passage.

Structural measurements accept Unicode writing. The current point-of-view and transition lists are English-specific evidence.

Element	What it records	What drift means
1. Sentence length	Mean words per sentence	The draft is materially shorter or longer than the reference.
2. Sentence variation	Spread of sentence lengths	The draft is unusually even or erratic.
3. Sentence structure	Frequent three-word opening shapes	Sentences begin in unfamiliar patterns.
4. Rhythm	Length change between neighbouring sentences	The beat has flattened or become needlessly choppy.
5. Paragraph length	Mean sentences per paragraph	Visual pacing has drifted.
6. Opening moves	Common first words	The piece enters unlike the reference samples.
7. Vocabulary	Frequent non-stop words	Familiar concrete language is missing or displaced.
8. Lexical density	Share of non-stop words	The draft is unusually generic or compressed.
9. Point of view	First-, second-, third-person, or mixed	Normal narrative distance changed.
10. Punctuation	Counts of selected marks	Emphasis or pauses differ from the reference.
11. Case style	Lowercase, standard, or mixed	Capitalization no longer matches the writer’s convention.
12. Question rate	Share of sentences ending in questions	The draft asks materially more or fewer questions.
13. Transitions	Frequent recognised connectors	The movement between ideas differs from the reference.

Current automated policy

• Explicit avoid-list phrases are red release blockers.
• Sentence-length drift is a yellow review cue outside max(8 words, 2.2 × profile variation) from the reference mean.
• Question-rate, case-style, and non-mixed point-of-view drift are yellow review cues.
• The remaining elements stay visible in the JSON profile and rewrite brief as evidence for an editor.

Yellow findings ask for editorial judgment. They do not prove a line is wrong. Any new automatic rule needs a written policy, positive tests, and counterexamples.