Prompt file imported from blockhead22/CRT-GroundCheck-SSE (.github/prompts/plan-phase12NegationDetection.prompt.md). Copyright stays with the author.

Phase 1.2: Add Negation Detection

Assessment: Phase 1.1 Progress

What's Working

1. Numeric mismatch check — Programming years 10→12 caught (Turn 9)
2. Entity swap detection — Employer Google contradiction caught (Turn 22)
3. Preference inversion — PhD→Master's denial caught (Turn 23)
4. Paraphrase gate tightening — No false positives

What's Still Missing (3 Cases)

Fix Plan

Priority 1: Add Negation Detection

File: personal_agent/crt_core.py Location: After _is_boolean_inversion method (around line 650)

The biggest gap. "I don't work at Google" vs "I work at Google" should trigger.

def _detect_negation_contradiction(self, text_new: str, text_prior: str) -> Tuple[bool, str]:
    """
    Detect negation-based contradictions.
    
    Patterns:
    - "I don't X" vs "I X"
    - "I no longer X" vs "I X"
    - "not X anymore" vs "X"
    """
    if not text_new or not text_prior:
        return False, ""
    
    text_new_lower = text_new.lower()
    text_prior_lower = text_prior.lower()
    
    # Negation patterns
    negation_patterns = [
        (r"(?:i\s+)?(?:don'?t|do\s+not|no\s+longer|not\s+anymore)\s+(\w+(?:\s+\w+){0,3})", "negated"),
        (r"(?:i\s+)?(?:stopped|quit|left|no\s+longer)\s+(\w+(?:\s+\w+){0,3})", "ceased"),
        (r"(?:i'm\s+not|i\s+am\s+not)\s+(\w+(?:\s+\w+){0,3})", "negated_state"),
    ]
    
    # Extract negated actions/states from new text
    negated_items = []
    for pattern, neg_type in negation_patterns:
        for match in re.finditer(pattern, text_new_lower):
            negated_items.append((match.group(1).strip(), neg_type))
    
    if not negated_items:
        return False, ""
    
    # Check if prior text affirms any of the negated items
    for item, neg_type in negated_items:
        # Clean item for matching
        item_words = item.split()[:3]  # First 3 words
        item_pattern = r'\b' + r'\s+'.join(re.escape(w) for w in item_words) + r'\b'
        
        # Check if prior affirms this (without negation)
        if re.search(item_pattern, text_prior_lower):
            # Verify prior doesn't also negate it
            prior_negated = any(
                re.search(p[0], text_prior_lower) 
                for p in negation_patterns
            )
            if not prior_negated:
                return True, f"Negation contradiction: '{item}' negated in new, affirmed in prior"
    
    return False, ""

Priority 2: Wire Negation Check into detect_contradiction

File: personal_agent/crt_core.py Location: In detect_contradiction method, after entity swap check (around line 520)

# ...existing code...
# Rule 0a: Entity swap detection
entity_swap, entity_reason = self._detect_entity_swap(slot, value_new, value_prior, text_new, text_prior)
if entity_swap:
    return True, entity_reason

# Rule 0b: Negation contradiction detection
negation_detected, negation_reason = self._detect_negation_contradiction(text_new, text_prior)
if negation_detected:
    return True, negation_reason

# Rule 0c: Preference/boolean inversion detection
# ...existing code...

Priority 3: Fix Fact Slot Wiring for Education

File: personal_agent/crt_rag.py Task: Verify education_level slot changes pass proper slot, value_new, value_prior parameters to contradiction detection

Execution Strategy

1. Apply Priority 1 & 2 — negation detection is the biggest gap
2. Re-run stress test to measure improvement
3. If still <80%, investigate fact extraction layer for Turn 7 (age) and Turn 24 (education)

Success Criteria

• Contradiction detection rate: ≥80% (target 100%)
• False positives: 0
• Caveat violations: ≤2
• Turn 21 negation ("don't work at Google") detected