Feature: Tricolon Detector

[craigtrim]

Overview

Tricolon (the rule of three) is one of the strongest AI stylistic tells. LLMs reach for three-part structure as the default skeleton for almost any explanatory or persuasive passage — with anomalously high frequency compared to human writers who use it deliberately and sparingly.

It operates at multiple levels:

Level	Example
Word/phrase	"Life, liberty, and the pursuit of happiness"
Clause	"It's not what you say, it's how you say it, or when you say it"
Sentence	"He came. He saw. He conquered."

Detection Approaches

Surface: Oxford comma tricolon

Regex match for X, Y, and Z or X, Y, or Z within a sentence.

Structural: Equal-segment comma split

Sentences containing exactly two commas where each comma-delimited segment is roughly equal in word count. No full syntactic parser needed — this is a pragmatic proxy for clausal tricolon.

Sentential: Three-sentence sequences

Consecutive sentences of similar syntactic class (sentence_types.py simple sentences) and similar length — detects "He came. He saw. He conquered." style tricolon.

Metrics

• tricolon_density: instances per 1000 words (the key signal — frequency, not presence)
• oxford_tricolon_count: surface-form list-of-three matches
• clausal_tricolon_count: equal-segment comma split matches
• sentential_tricolon_count: consecutive parallel-sentence triples

Related

• Feature: AI Stylistic Tell Detection — Tricolon, Stacked Sentences, and AI Vocabulary Fingerprints #69 — AI Stylistic Tell Detection (parent feature)
• Feature: Tricolon Detector #73 — Co-occurrence scorer (downstream consumer)

[craigtrim]

Implementation Notes

Forms 1 and 2 can ship independently of #72

Form 1 — Oxford Tricolon (X, Y, and/or Z):

• Split sentence by comma, check for 3 segments where the last starts with "and" or "or"
• Edge case: nested commas inside items break simple splitting — if any segment itself contains a comma, flag as lower-confidence rather than discarding
• ~80% of cases handled cleanly with the simple split

Form 2 — Clausal Equal-Segment (catches asyndeton too: "I came, I saw, I conquered"):

• Find sentences with exactly 2 commas, split into 3 segments
• Compute CV of segment word counts — CV below ~0.35 indicates parallel structure
• Works because parallel clauses instantiate the same syntactic template and tend toward similar length; random two-comma sentences (subordinate clauses, parentheticals) have much higher segment variance

Form 3 is blocked by #72

Form 3 — Sentential Tricolon ("He came. He saw. He conquered."):

• Requires paragraph boundaries to identify three-sentence sequences reliably — without them you cannot distinguish a terminal three-sentence sequence from sentences spanning a paragraph break
• Approach: within each paragraph, sliding window of 3 consecutive sentences where all are classified simple by sentence_types.py (already exists), all ≤ ~8 words, and CV of lengths is low
• Feature: Paragraph-Level Segmentation #72 (Paragraph-Level Segmentation) is a hard blocker for this form

Recommended Build Order

1. Ship forms 1 and 2 without Feature: Paragraph-Level Segmentation #72
2. Add form 3 once Feature: Paragraph-Level Segmentation #72 is complete

On False Positives

Forms 1 and 2 will produce false positives (e.g. "In 2024, researchers found that, on average, results improved" passes the clausal test). This is acceptable because density is the signal, not presence — noise from individual false positives doesn't move the aggregate score at scale. An author with 8 tricolons in 500 words is the target, not the single ambiguous case.

Output Shape

TricolonResult:
  oxford_count          # Form 1 matches
  clausal_count         # Form 2 matches
  sentential_count      # Form 3 matches (requires #72)
  tricolon_density      # total instances per 1000 words
  tricolon_paragraph_rate   # proportion of paragraphs containing ≥1 tricolon
  instances             # list of {form, sentence_text, position}

[craigtrim]

Blockers

All of the following must be completed before tricolon candidate validation is possible:

#	Title	Gap it fills
#72	Paragraph-Level Segmentation	Required for sentential (Form 3) tricolon detection
#76	Beat Detection	Required for #78 cross-item stress shape comparison
#77	Phrase-Level Syllabic Homogeneity	Rejects candidates with outlier syllable weights
#78	Cross-Item Stress Shape Comparator	Classifies isocolon / climactic / irregular beat shape
#79	Item-Level Alliteration Detector	Boosts confidence for alliterative candidates
#80	Head-Word POS Extractor	Verifies syntactic parallelism across items

Once all six are complete, tricolon detection reduces to:

1. Regex candidate gate (Forms 1 and 2 immediately, Form 3 after Feature: Paragraph-Level Segmentation #72)
2. Score each candidate through Feature: Phrase-Level Syllabic Homogeneity Scorer #77 + Feature: Cross-Item Stress Shape Comparator #78 + Feature: Item-Level Alliteration Detector for Candidate Phrases #79 + Feature: Head-Word POS Extractor for Phrase Strings #80
3. Classify confidence band (Strong / Moderate / Weak / Rejected)
4. Aggregate density metric