feat: self-contradiction detection via negation polarity

[Jamie-BitFlight]

Summary

Detect self-contradictory statements within a single assistant response using negation polarity heuristics. When the assistant says "X is Y" in one sentence and "X is not Y" in another, flag it as a contradiction.

Technique (from ObvioSpectre/hallucination-detector)

Lightweight NLI approximation without requiring an NLI model:

1. Split response into sentences
2. For each sentence pair on the same topic:
   • Check if one contains negation words and the other does not
   • If so, flag as internal contradiction

Negation Words

const NEGATION_WORDS = ['not', 'no', 'never', "didn't", "isn't", "wasn't", "aren't", "won't", "can't", "doesn't", 'none', 'neither', 'nor', 'unable', 'lacks', 'failed'];

Detection Logic (Regex-Adaptable)

Without embeddings, approximate "same topic" by checking for shared noun phrases or subjects:

// Pattern: "X is Y" ... "X is not Y"
// Pattern: "always X" ... "never X"  
// Pattern: "X works" ... "X doesn't work"

Why This Works

Self-contradiction is a strong hallucination signal — the assistant is confabulating rather than reasoning from consistent evidence. This catches a failure mode that none of our current four categories detect.

New Category

internal_contradiction — a sixth detection category.

Acceptance Criteria

• Sentence splitting implemented (regex: split on .!? followed by whitespace)
• Negation polarity detection for sentence pairs
• Handles common negation patterns beyond simple "not"
• Low false positive rate — only flags when sentences are on the same topic
• Tests with clear contradiction examples and non-contradiction controls
• Suppression for quoted text and code blocks

References

• ObvioSpectre/hallucination-detector detectors/consistency.py
• Negation-polarity heuristic approximates NLI contradiction detection

[Jamie-BitFlight]

Grouping: Group C — Sentence-Level Detection

Implement together with: #4, #20

Rationale: #19 and #4 both detect self-contradiction — #4 via proximity-based contradiction pairs, #19 via negation polarity. These should be combined into one self_contradiction category using both techniques as complementary signals rather than two separate categories.

All three Group C issues share the sentence-splitting infrastructure.

Dependency: The sentence splitter built here becomes the foundation for #21 (sentence-level scoring).

[Jamie-BitFlight]

Deep Analysis: Implementation Impact

Classification

• Impact type: STRUCTURAL — introduces sentence splitting stage
• Risk: Medium — new processing pipeline stage
• Phase: 3 — Sentence infrastructure

Introduces sentence splitting

The sentence splitter (`text.split(/(?<=[.!?])\s+/)`) is a new processing stage. This regex has known edge cases:

• Abbreviations: "Dr. Smith" → false split
• Decimals: "version 3.2. Next" → false split
• URLs: "https://example.com. Then" → false split
• Ellipses: "wait... then" → false split

These edge cases need suppression rules or a more robust splitter.

Combine with #4

Both #4 and #19 detect self-contradiction. Combine into one `self_contradiction` category:

• feat: add self-contradiction detection #4's proximity pairs catch same-paragraph contradictions
• feat: self-contradiction detection via negation polarity #19's negation polarity catches cross-sentence contradictions

Shared infrastructure

The sentence splitter becomes shared infrastructure for #4, #20, and #21. Extract as a reusable function (`splitSentences()`).

Sequencing

• Blocked by: nothing
• Implement together with: feat: add self-contradiction detection #4 (combine), feat: degenerate repetition / collapse detection #20 (shared splitter)
• Provides infrastructure for: feat: sentence-level granularity and weighted multi-signal scoring #21 (sentence-level scoring)

State: None | Dependencies: None | Contract change: None (new kind only)