Overview
Given a list of text phrases (candidate parallel units), determine whether their first content words share an initial phoneme — and whether that constitutes full alliteration or partial assonance. This is the alliteration validation layer for tricolon candidate scoring.
Why It's Needed
Alliterative tricolon is one of the highest-confidence detections: "wild, wanton, and wicked" — the shared initial phoneme across all three items is near-proof of intentional parallelism. It dramatically reduces false positives and increases candidate confidence score.
Beyond tricolon, alliteration is an AI tell in its own right — LLMs generate alliterative lists because they've learned it signals "polished writing." High alliterative tricolon density is a compounding signal.
What Already Exists
In prosody/rhythm_prosody.py:
_get_initial_sound(word) — returns initial consonant phoneme from CMU dict, falls back to first letter. This is exactly what's needed._compute_alliteration(words) — counts alliterative consecutive word pairs per 100 words across a full document.
The gap: _compute_alliteration operates across a flat word list at document scale. It cannot answer "do the lead words of these three specific phrases alliterate with each other?" — which is what tricolon validation requires.
Proposed Implementation
def detect_item_alliteration(phrases: list[str]) -> ItemAlliterationResult:
# Extract first content word of each phrase
# Get initial phoneme via _get_initial_sound()
# Compare across all items
# full_alliteration: all items share the same initial phoneme
# partial_alliteration: majority share
# assonance: shared vowel onset rather than consonantOutput
ItemAlliterationResult:
lead_phonemes: list[str | None] # initial phoneme per phrase
full_alliteration: bool # all items match
partial_alliteration: bool # majority match
assonance: bool # shared vowel sound (not consonant)
alliteration_score: float # 0-1, proportion of items that match lead
Acceptance Criteria
["wild", "wanton", "wicked"] → full_alliteration=True, alliteration_score=1.0["quick", "smart", "brave"] → full_alliteration=False, alliteration_score=0.0["came", "saw", "conquered"] → full_alliteration=False (K / S / K — partial)
Blockers / Downstream
Related
Overview
Given a list of text phrases (candidate parallel units), determine whether their first content words share an initial phoneme — and whether that constitutes full alliteration or partial assonance. This is the alliteration validation layer for tricolon candidate scoring.
Why It's Needed
Alliterative tricolon is one of the highest-confidence detections:
"wild, wanton, and wicked"— the shared initial phoneme across all three items is near-proof of intentional parallelism. It dramatically reduces false positives and increases candidate confidence score.Beyond tricolon, alliteration is an AI tell in its own right — LLMs generate alliterative lists because they've learned it signals "polished writing." High alliterative tricolon density is a compounding signal.
What Already Exists
In
prosody/rhythm_prosody.py:_get_initial_sound(word)— returns initial consonant phoneme from CMU dict, falls back to first letter. This is exactly what's needed._compute_alliteration(words)— counts alliterative consecutive word pairs per 100 words across a full document.The gap:
_compute_alliterationoperates across a flat word list at document scale. It cannot answer "do the lead words of these three specific phrases alliterate with each other?" — which is what tricolon validation requires.Proposed Implementation
Output
Acceptance Criteria
["wild", "wanton", "wicked"]→full_alliteration=True,alliteration_score=1.0["quick", "smart", "brave"]→full_alliteration=False,alliteration_score=0.0["came", "saw", "conquered"]→full_alliteration=False(K / S / K — partial)Blockers / Downstream
_get_initial_soundalready exists, can be built immediatelyRelated
prosody/rhythm_prosody.py—_get_initial_soundis the key building block