Feature: AI-Tell Co-occurrence Scorer

[craigtrim]

Overview

Individual AI stylistic tells are weak signals in isolation — a human writer might occasionally use tricolon, or end a paragraph with a short sentence. The strength of the detection is in co-occurrence: multiple tells appearing together repeatedly across a document.

This module aggregates existing and new per-feature scores into a single weighted confidence score.

Input Signals

From existing modules (no new code needed):

Signal	Source	Module
complexity_uniformity_score	Syllabic uniformity	prosody/sentence_syllable_patterns.py
repetition_ratio	Formulaic n-gram patterns	prosody/syllable_pattern_repetition.py
starting_pattern_repetition_rate	Formulaic openings	prosody/syllable_pattern_repetition.py
ending_pattern_repetition_rate	Formulaic closings	prosody/syllable_pattern_repetition.py
pattern_entropy	Distribution concentration	prosody/syllable_pattern_repetition.py

From new modules (pending issues):

Signal	Source	Issue
tricolon_density	Rule of three frequency	#70
terminal_brevity_ratio	Mic drop paragraph shape	#71 (paragraph segmentation)
short_paragraph_run_length	Stacked short paragraphs	#71 (paragraph segmentation)
ai_vocabulary_score	LLM-preferred word frequency	#72

Composite Score

ai_tell_score = weighted_mean([
    complexity_uniformity_score,
    repetition_ratio,
    tricolon_density_normalized,
    terminal_brevity_score,
    stacked_paragraph_score,
    ai_vocabulary_score,
])

Weights should be empirically tunable, with sensible defaults derived from signal reliability.

Output

• ai_tell_score: 0.0–1.0 aggregate confidence
• signal_breakdown: per-feature contribution scores
• dominant_signal: which feature contributed most
• Interpretive band: Low / Moderate / High / Very High

Relationship to #68

The co-occurrence score feeds directly into the style conformance penalty in #68 (Tonality Derivation). High ai_tell_score in LLM-generated output is a direct deduction against fidelity to a human author's tonality.

Related

• Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69 — AI Stylistic Tell Detection (parent feature)
• Feature: Tonality Derivation and Generative AI Style Transfer Validation #68 — Tonality Derivation and Style Transfer Validation
• Feature: Add AI-Generated Text Detection Module #65 — AI-Generated Text Detection
• Feature: Tricolon Detector — Rule of Three Pattern Recognition #70 — Tricolon Detector
• Feature: Paragraph-Level Segmentation #71 — Paragraph-Level Segmentation
• Feature: Paragraph-Level Segmentation #72 — AI Vocabulary Wordlist

[craigtrim]

Paragraph Segmentation Blocker Is Resolved

The two paragraph-dependent input signals (terminal_brevity_ratio, short_paragraph_run_length) are unblocked. fast-sentence-segment's segment() already returns list[list[str]] — paragraphs as outer lists, sentences as inner. See #71 for full analysis.

Once split_paragraphs() is added to _utils.py (thin wrapper, no new dependency — see #71), both signals can be computed:

Signal	How it's derived from split_paragraphs() output
terminal_brevity_ratio	last_sentence_word_count / mean(paragraph_sentence_word_counts) — per paragraph
short_paragraph_run_length	Longest consecutive run of single-sentence paragraphs
single_sentence_paragraph_ratio	count(paragraphs with len==1) / total_paragraphs

Blocking Order

1. split_paragraphs() in _utils.py — Feature: Paragraph-Level Segmentation #71 (closes immediately, no new work)
2. Mic drop + stacked paragraph detectors — Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69
3. This issue (Feature: AI-Tell Co-occurrence Scorer #75) — aggregates outputs from Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69 and Feature: Tricolon Detector — Rule of Three Pattern Recognition #70 and Feature: Tricolon Detector #73

Related

• Feature: Paragraph-Level Segmentation #71 — adds split_paragraphs() (immediate prerequisite, can close now)
• Feature: Paragraph-Level Segmentation #72 — duplicate of Feature: Paragraph-Level Segmentation #71
• Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69 — paragraph-dependent detectors that feed this scorer
• Feature: Tricolon Detector — Rule of Three Pattern Recognition #70 — tricolon detector (also feeds this scorer)
• Feature: Tricolon Detector #73 — tricolon detector (also feeds this scorer)
• craigtrim/fast-sentence-segment — source library with native paragraph output