explanation_math_firewall.py

# explanation_math_firewall.py - V4.2 (Behavioral Firewall)
# Prevents semantic leaks of forbidden mathematical concepts.

import re
import logging

logger = logging.getLogger(__name__)

# Concepts that are often "shadow paths" or over-qualified for lower grades
FORBIDDEN_CONCEPTS = [
    r"נגזרת", r"נגזור", r"גזירה", r"קיצון", r"מקסימום", r"מינימום", 
    r"אינטגרל", r"אינטגרציה", r"הפרש", r"מערכת", r"נעלמים", r"שני נעלמים",
    r"פיתגורס", r"שורשים", r"מכנה משותף" # Add more as needed
]

# Symbols that indicate over-qualified math
FORBIDDEN_SYMBOLS = [
    r"'", r"\\'", r"integral", r"\\int", r"\\Sigma", r"\\Delta"
]

def scan_explanation(text: str, proof_graph) -> tuple[bool, str]:
    """
    Scans explanation text for concepts NOT present in the ProofGraph.
    Returns (is_safe, violation_reason).
    """
    if not text:
        return True, ""

    # 1. Check for Hebrew forbidden concepts
    for pattern in FORBIDDEN_CONCEPTS:
        if re.search(pattern, text):
            # Check if this concept is "allowed" because it's in the ProofGraph
            # (e.g., if we actually used a Derivative, it's fine to say 'נגזרת')
            is_in_proof = False
            if proof_graph:
                for step in proof_graph.steps:
                    if step.operator_used and pattern in step.logic_description:
                        is_in_proof = True
                        break
            
            if not is_in_proof:
                logger.warning(f"🛡️ [FIREWALL] Semantic violation detected: '{pattern}'")
                return False, f"Concept '{pattern}' found in text but not in ProofGraph."

    # 2. Check for forbidden mathematical symbols
    for symbol in FORBIDDEN_SYMBOLS:
        if symbol in text:
            # Similar logic: check if the symbol appears in math_content of proof_graph
            is_in_proof = False
            if proof_graph:
                for step in proof_graph.steps:
                    if symbol in str(step.math_content):
                        is_in_proof = True
                        break
            
            if not is_in_proof:
                logger.warning(f"🛡️ [FIREWALL] Symbol violation detected: '{symbol}'")
                return False, f"Symbol '{symbol}' found in text but not in ProofGraph."

    return True, ""