agent_os_kernel 3.1.0__py3-none-any.whl

cmvk/verification.py

@@ -0,0 +1,956 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+"""
+CMVK Verification Module - Pure Mathematical Functions
+
+This module provides pure functions for calculating drift/hallucination scores
+between two outputs. These functions have no side effects and use only
+numpy/scipy for mathematical operations.
+
+Layer 1: The Primitive - Mathematical and adversarial verification.
+
+Enhanced Features (v0.2.0):
+    - Configurable distance metrics (cosine, euclidean, manhattan, etc.)
+    - Dimensional weighting for importance-based drift calculation
+    - Threshold profiles for domain-specific verification
+    - Explainable drift with per-dimension contributions
+    - Batch verification for efficiency
+    - Audit trail integration
+"""
+
+from __future__ import annotations
+
+from collections.abc import Sequence
+from dataclasses import dataclass
+from enum import Enum
+from typing import TYPE_CHECKING, Any
+
+import numpy as np
+from numpy.typing import ArrayLike
+
+try:
+    from scipy import stats
+
+    HAS_SCIPY = True
+except ImportError:
+    HAS_SCIPY = False
+
+if TYPE_CHECKING:
+    from .audit import AuditTrail
+
+
+class DriftType(Enum):
+    """Types of drift/divergence detected between outputs."""
+
+    SEMANTIC = "semantic"
+    STRUCTURAL = "structural"
+    NUMERICAL = "numerical"
+    LEXICAL = "lexical"
+
+
+@dataclass(frozen=True)
+class VerificationScore:
+    """
+    Immutable result of verification between two outputs.
+
+    Attributes:
+        drift_score: Overall drift score between 0.0 (identical) and 1.0 (completely different)
+        confidence: Confidence in the score (0.0 to 1.0)
+        drift_type: Primary type of drift detected
+        details: Dictionary with component scores
+        explanation: Optional drift explanation with dimension contributions (CMVK-010)
+    """
+
+    drift_score: float
+    confidence: float
+    drift_type: DriftType
+    details: dict
+    explanation: dict | None = None
+
+    def passed(self, threshold: float = 0.3) -> bool:
+        """Check if drift is within acceptable threshold."""
+        return self.drift_score <= threshold
+
+    def to_dict(self) -> dict:
+        """Convert to dictionary for serialization."""
+        return {
+            "drift_score": self.drift_score,
+            "confidence": self.confidence,
+            "drift_type": self.drift_type.value,
+            "details": self.details,
+            "explanation": self.explanation,
+        }
+
+
+@dataclass(frozen=True)
+class DriftExplanation:
+    """
+    Detailed explanation of drift between two vectors (CMVK-010).
+
+    Attributes:
+        primary_drift_dimension: Index or name of dimension with highest contribution
+        dimension_contributions: Mapping of dimension to its contribution percentage
+        top_contributors: List of top N contributing dimensions
+        metric_used: The distance metric used
+        interpretation: Human-readable interpretation of the drift
+    """
+
+    primary_drift_dimension: str | int
+    dimension_contributions: dict[str | int, float]
+    top_contributors: list[tuple[str | int, float]]
+    metric_used: str
+    interpretation: str
+
+    def to_dict(self) -> dict:
+        """Convert to dictionary."""
+        return {
+            "primary_drift_dimension": self.primary_drift_dimension,
+            "dimension_contributions": self.dimension_contributions,
+            "top_contributors": self.top_contributors,
+            "metric_used": self.metric_used,
+            "interpretation": self.interpretation,
+        }
+
+
+def verify(output_a: str, output_b: str) -> VerificationScore:
+    """
+    Calculate drift/hallucination score between two outputs.
+
+    This is the primary verification function - a pure function with no side effects.
+    Takes two outputs and returns a score indicating their divergence.
+
+    Args:
+        output_a: First output (typically from model A / generator)
+        output_b: Second output (typically from model B / verifier)
+
+    Returns:
+        VerificationScore with drift score, confidence, and details
+
+    Example:
+        >>> score = verify("def add(a, b): return a + b", "def add(x, y): return x + y")
+        >>> score.drift_score  # Low score - semantically similar
+        0.15
+    """
+    if not output_a and not output_b:
+        return VerificationScore(
+            drift_score=0.0,
+            confidence=1.0,
+            drift_type=DriftType.LEXICAL,
+            details={"reason": "both_empty"},
+        )
+
+    if not output_a or not output_b:
+        return VerificationScore(
+            drift_score=1.0,
+            confidence=1.0,
+            drift_type=DriftType.STRUCTURAL,
+            details={"reason": "one_empty"},
+        )
+
+    # Calculate multiple drift components
+    lexical_drift = _lexical_drift(output_a, output_b)
+    structural_drift = _structural_drift(output_a, output_b)
+    numerical_drift = _numerical_drift(output_a, output_b)
+
+    # Weighted combination
+    weights = {"lexical": 0.3, "structural": 0.4, "numerical": 0.3}
+
+    combined_drift = (
+        weights["lexical"] * lexical_drift["score"]
+        + weights["structural"] * structural_drift["score"]
+        + weights["numerical"] * numerical_drift["score"]
+    )
+
+    # Determine primary drift type
+    scores = {
+        DriftType.LEXICAL: lexical_drift["score"],
+        DriftType.STRUCTURAL: structural_drift["score"],
+        DriftType.NUMERICAL: numerical_drift["score"],
+    }
+    primary_drift = max(scores, key=lambda k: scores[k])
+
+    # Calculate confidence based on agreement between methods
+    score_values = list(scores.values())
+    confidence = 1.0 - np.std(score_values) if len(score_values) > 1 else 0.8
+
+    return VerificationScore(
+        drift_score=float(np.clip(combined_drift, 0.0, 1.0)),
+        confidence=float(np.clip(confidence, 0.0, 1.0)),
+        drift_type=primary_drift,
+        details={
+            "lexical": lexical_drift,
+            "structural": structural_drift,
+            "numerical": numerical_drift,
+            "weights": weights,
+        },
+    )
+
+
+def verify_embeddings(
+    embedding_a: ArrayLike,
+    embedding_b: ArrayLike,
+    metric: str = "cosine",
+    weights: ArrayLike | None = None,
+    threshold_profile: str | None = None,
+    explain: bool = False,
+    dimension_names: list[str] | None = None,
+    audit_trail: AuditTrail | None = None,
+) -> VerificationScore:
+    """
+    Calculate drift score between two embedding vectors.
+
+    Enhanced verification function with configurable metrics, weighting,
+    threshold profiles, and explainability (CMVK-001 through CMVK-010).
+
+    Args:
+        embedding_a: Embedding vector for output A (e.g., claimed values)
+        embedding_b: Embedding vector for output B (e.g., observed values)
+        metric: Distance metric to use. Options:
+            - "cosine": Cosine distance (default, normalizes vectors)
+            - "euclidean": Euclidean distance (preserves magnitude - CMVK-001)
+            - "manhattan": Manhattan/L1 distance
+            - "chebyshev": Maximum absolute difference
+            - "mahalanobis": Mahalanobis distance
+        weights: Optional weights for each dimension (CMVK-008).
+                Higher weights increase that dimension's contribution to drift.
+        threshold_profile: Name of threshold profile to use (CMVK-005).
+                         Options: "carbon", "financial", "medical", "general", "strict"
+        explain: If True, include detailed drift explanation (CMVK-010)
+        dimension_names: Optional names for dimensions (for explainability)
+        audit_trail: Optional AuditTrail instance for logging (CMVK-006)
+
+    Returns:
+        VerificationScore with drift score, confidence, and optional explanation
+
+    Example:
+        >>> # Basic usage
+        >>> score = verify_embeddings(claim_vec, obs_vec)
+
+        >>> # With Euclidean distance for magnitude-sensitive comparison
+        >>> score = verify_embeddings(
+        ...     claim_vec, obs_vec,
+        ...     metric="euclidean",
+        ...     threshold_profile="carbon",
+        ...     explain=True
+        ... )
+
+        >>> # With dimensional weighting
+        >>> score = verify_embeddings(
+        ...     claim_vec, obs_vec,
+        ...     metric="euclidean",
+        ...     weights=[0.6, 0.4],  # NDVI more important than carbon
+        ...     explain=True,
+        ...     dimension_names=["ndvi", "carbon_stock"]
+        ... )
+    """
+    from .metrics import calculate_distance, calculate_weighted_distance
+
+    vec_a = np.asarray(embedding_a, dtype=np.float64)
+    vec_b = np.asarray(embedding_b, dtype=np.float64)
+
+    # Load threshold profile if specified
+    profile = None
+    if threshold_profile:
+        from .profiles import get_profile
+
+        profile = get_profile(threshold_profile)
+        # Use profile's default metric if none specified
+        if metric == "cosine" and profile.default_metric != "cosine":
+            metric = profile.default_metric
+
+    # Shape validation
+    if vec_a.shape != vec_b.shape:
+        result = VerificationScore(
+            drift_score=1.0,
+            confidence=0.5,
+            drift_type=DriftType.STRUCTURAL,
+            details={"reason": "shape_mismatch", "shape_a": vec_a.shape, "shape_b": vec_b.shape},
+        )
+        if audit_trail:
+            _log_to_audit(audit_trail, vec_a, vec_b, result, metric, threshold_profile)
+        return result
+
+    # Calculate distance with appropriate function
+    if weights is not None:
+        metric_result = calculate_weighted_distance(vec_a, vec_b, weights=weights, metric=metric)
+    else:
+        metric_result = calculate_distance(vec_a, vec_b, metric=metric)
+
+    # Build drift score from normalized distance
+    drift_score = float(np.clip(metric_result.normalized, 0.0, 1.0))
+
+    # Calculate confidence based on vector properties
+    confidence = _calculate_embedding_confidence(vec_a, vec_b)
+
+    # Build explanation if requested
+    explanation_dict = None
+    if explain:
+        explanation = _build_drift_explanation(
+            vec_a, vec_b, metric_result, weights, dimension_names
+        )
+        explanation_dict = explanation.to_dict()
+
+    # Build details
+    details = {
+        "metric": metric,
+        "raw_distance": metric_result.distance,
+        "normalized_distance": metric_result.normalized,
+        **metric_result.details,
+    }
+
+    # Add profile info if used
+    if profile:
+        passed = profile.is_within_threshold(drift_score, confidence)
+        severity = profile.get_severity(drift_score)
+        details["profile"] = {
+            "name": profile.name,
+            "drift_threshold": profile.drift_threshold,
+            "passed": passed,
+            "severity": severity,
+        }
+
+    result = VerificationScore(
+        drift_score=drift_score,
+        confidence=confidence,
+        drift_type=DriftType.SEMANTIC,
+        details=details,
+        explanation=explanation_dict,
+    )
+
+    # Log to audit trail if provided
+    if audit_trail:
+        _log_to_audit(audit_trail, vec_a, vec_b, result, metric, threshold_profile)
+
+    return result
+
+
+def verify_embeddings_batch(
+    embeddings_a: Sequence[ArrayLike],
+    embeddings_b: Sequence[ArrayLike],
+    metric: str = "cosine",
+    weights: ArrayLike | None = None,
+    threshold_profile: str | None = None,
+    explain: bool = False,
+    dimension_names: list[str] | None = None,
+    audit_trail: AuditTrail | None = None,
+) -> list[VerificationScore]:
+    """
+    Verify multiple embedding pairs efficiently (CMVK-004).
+
+    Processes all pairs with consistent settings and optional audit logging.
+
+    Args:
+        embeddings_a: Sequence of embedding vectors from source A
+        embeddings_b: Sequence of embedding vectors from source B
+        metric: Distance metric (applied to all pairs)
+        weights: Dimensional weights (applied to all pairs)
+        threshold_profile: Threshold profile name
+        explain: Whether to include explanations
+        dimension_names: Optional dimension names for explainability
+        audit_trail: Optional AuditTrail for logging
+
+    Returns:
+        List of VerificationScore for each pair
+
+    Raises:
+        ValueError: If sequence lengths don't match
+    """
+    if len(embeddings_a) != len(embeddings_b):
+        raise ValueError(
+            f"Length mismatch: embeddings_a has {len(embeddings_a)} items, "
+            f"embeddings_b has {len(embeddings_b)} items"
+        )
+
+    results = []
+    for vec_a, vec_b in zip(embeddings_a, embeddings_b, strict=True):
+        score = verify_embeddings(
+            vec_a,
+            vec_b,
+            metric=metric,
+            weights=weights,
+            threshold_profile=threshold_profile,
+            explain=explain,
+            dimension_names=dimension_names,
+            audit_trail=audit_trail,
+        )
+        results.append(score)
+
+    return results
+
+
+def aggregate_embedding_scores(
+    scores: Sequence[VerificationScore],
+    threshold_profile: str | None = None,
+) -> dict[str, Any]:
+    """
+    Aggregate multiple embedding verification scores with profile context.
+
+    Args:
+        scores: Sequence of VerificationScore objects
+        threshold_profile: Optional profile for pass/fail classification
+
+    Returns:
+        Dictionary with aggregate statistics and pass rates
+    """
+    if not scores:
+        return {"count": 0}
+
+    profile = None
+    if threshold_profile:
+        from .profiles import get_profile
+
+        profile = get_profile(threshold_profile)
+
+    drift_values = [s.drift_score for s in scores]
+    confidence_values = [s.confidence for s in scores]
+
+    # Calculate pass/fail if profile available
+    if profile:
+        passed_count = sum(
+            1 for s in scores if profile.is_within_threshold(s.drift_score, s.confidence)
+        )
+        severity_counts: dict[str, int] = {
+            "pass": 0,
+            "warning": 0,
+            "critical": 0,
+            "severe": 0,
+        }
+        for s in scores:
+            severity = profile.get_severity(s.drift_score)
+            severity_counts[severity] += 1
+    else:
+        passed_count = sum(1 for s in scores if s.drift_score <= 0.3)
+        severity_counts = {}
+
+    result: dict[str, Any] = {
+        "count": len(scores),
+        "passed_count": passed_count,
+        "failed_count": len(scores) - passed_count,
+        "pass_rate": passed_count / len(scores),
+        "mean_drift": float(np.mean(drift_values)),
+        "std_drift": float(np.std(drift_values)),
+        "min_drift": float(np.min(drift_values)),
+        "max_drift": float(np.max(drift_values)),
+        "median_drift": float(np.median(drift_values)),
+        "mean_confidence": float(np.mean(confidence_values)),
+        "p95_drift": float(np.percentile(drift_values, 95)),
+    }
+
+    if severity_counts and profile:
+        result["severity_distribution"] = severity_counts
+        result["profile_used"] = profile.name
+
+    return result
+
+
+# ============================================================================
+# Explainability Functions (CMVK-010)
+# ============================================================================
+
+
+def _build_drift_explanation(
+    vec_a: np.ndarray,
+    vec_b: np.ndarray,
+    metric_result: Any,
+    weights: ArrayLike | None,
+    dimension_names: list[str] | None,
+) -> DriftExplanation:
+    """Build detailed drift explanation."""
+    diff = np.abs(vec_a - vec_b)
+
+    # Apply weights if provided
+    if weights is not None:
+        weights_arr = np.asarray(weights, dtype=np.float64)
+        weighted_diff = diff * weights_arr
+    else:
+        weighted_diff = diff
+
+    # Calculate per-dimension contributions
+    total_diff = np.sum(weighted_diff)
+    contributions = weighted_diff / total_diff if total_diff > 0 else np.zeros_like(diff)
+
+    # Map contributions to names or indices
+    contrib_dict: dict[str | int, float]
+    sorted_contribs: list[tuple[str | int, float]]
+    primary_dim: str | int
+
+    if dimension_names and len(dimension_names) == len(contributions):
+        contrib_dict = {
+            name: float(c) for name, c in zip(dimension_names, contributions, strict=False)
+        }
+        sorted_contribs = sorted(contrib_dict.items(), key=lambda x: x[1], reverse=True)
+        primary_dim = sorted_contribs[0][0]
+    else:
+        contrib_dict = {i: float(c) for i, c in enumerate(contributions)}
+        sorted_contribs = sorted(contrib_dict.items(), key=lambda x: x[1], reverse=True)
+        primary_dim = sorted_contribs[0][0]
+
+    # Top contributors (up to 5)
+    top_contributors: list[tuple[str | int, float]] = sorted_contribs[:5]
+
+    # Generate interpretation
+    interpretation = _generate_interpretation(
+        vec_a, vec_b, primary_dim, top_contributors, dimension_names
+    )
+
+    return DriftExplanation(
+        primary_drift_dimension=primary_dim,
+        dimension_contributions=contrib_dict,
+        top_contributors=top_contributors,
+        metric_used=metric_result.metric.value,
+        interpretation=interpretation,
+    )
+
+
+def _generate_interpretation(
+    vec_a: np.ndarray,
+    vec_b: np.ndarray,
+    primary_dim: str | int,
+    top_contributors: list[tuple[str | int, float]],
+    dimension_names: list[str] | None,
+) -> str:
+    """Generate human-readable interpretation of drift."""
+    # Get primary dimension index
+    if isinstance(primary_dim, str) and dimension_names:
+        idx = dimension_names.index(primary_dim)
+    else:
+        idx = primary_dim if isinstance(primary_dim, int) else 0
+
+    diff_value = abs(vec_a[idx] - vec_b[idx])
+    pct_diff = (diff_value / abs(vec_a[idx])) * 100 if vec_a[idx] != 0 else float("inf")
+
+    dim_name = primary_dim if isinstance(primary_dim, str) else f"dimension {primary_dim}"
+
+    if len(top_contributors) > 1 and top_contributors[0][1] > 0.5:
+        return (
+            f"Drift primarily driven by {dim_name} "
+            f"({top_contributors[0][1]*100:.1f}% of total drift). "
+            f"Value changed from {vec_a[idx]:.4f} to {vec_b[idx]:.4f} "
+            f"({pct_diff:.1f}% difference)."
+        )
+    elif len(top_contributors) > 1:
+        top_names = [str(c[0]) for c, _ in zip(top_contributors[:3], range(3), strict=False)]
+        return (
+            f"Drift distributed across multiple dimensions. "
+            f"Top contributors: {', '.join(top_names)}. "
+            f"Largest single change in {dim_name}."
+        )
+    else:
+        return f"Single dimension drift in {dim_name}."
+
+
+def _calculate_embedding_confidence(
+    vec_a: np.ndarray,
+    vec_b: np.ndarray,
+) -> float:
+    """Calculate confidence score for embedding verification."""
+    # Base confidence
+    confidence = 0.9
+
+    # Reduce confidence for very small vectors (less reliable)
+    if len(vec_a) < 10:
+        confidence *= 0.9
+
+    # Reduce confidence if vectors have very different magnitudes
+    norm_a = np.linalg.norm(vec_a)
+    norm_b = np.linalg.norm(vec_b)
+    if norm_a > 0 and norm_b > 0:
+        magnitude_ratio = min(norm_a, norm_b) / max(norm_a, norm_b)
+        if magnitude_ratio < 0.5:
+            confidence *= 0.85
+
+    # Reduce confidence for near-zero vectors
+    if norm_a < 1e-6 or norm_b < 1e-6:
+        confidence *= 0.7
+
+    return float(np.clip(confidence, 0.0, 1.0))
+
+
+def _log_to_audit(
+    audit_trail: AuditTrail,
+    vec_a: np.ndarray,
+    vec_b: np.ndarray,
+    result: VerificationScore,
+    metric: str,
+    profile_name: str | None,
+) -> None:
+    """Log verification to audit trail."""
+    passed = result.details.get("profile", {}).get("passed", result.drift_score <= 0.3)
+
+    audit_trail.log(
+        operation="verify_embeddings",
+        inputs={
+            "embedding_a_shape": vec_a.shape,
+            "embedding_b_shape": vec_b.shape,
+            "embedding_a_norm": float(np.linalg.norm(vec_a)),
+            "embedding_b_norm": float(np.linalg.norm(vec_b)),
+        },
+        drift_score=result.drift_score,
+        confidence=result.confidence,
+        metric_used=metric,
+        profile_used=profile_name,
+        passed=passed,
+        result_details={
+            "drift_type": result.drift_type.value,
+            "raw_distance": result.details.get("raw_distance"),
+        },
+    )
+
+
+def verify_distributions(dist_a: ArrayLike, dist_b: ArrayLike) -> VerificationScore:
+    """
+    Calculate drift between two probability distributions.
+
+    Uses KL divergence and other statistical measures to compare distributions.
+
+    Args:
+        dist_a: First probability distribution
+        dist_b: Second probability distribution
+
+    Returns:
+        VerificationScore with distribution-based drift score
+    """
+    p = np.asarray(dist_a, dtype=np.float64)
+    q = np.asarray(dist_b, dtype=np.float64)
+
+    # Normalize to valid probability distributions
+    p = p / (p.sum() + 1e-10)
+    q = q / (q.sum() + 1e-10)
+
+    # Add small epsilon to avoid log(0)
+    eps = 1e-10
+    p = np.clip(p, eps, 1.0)
+    q = np.clip(q, eps, 1.0)
+
+    if HAS_SCIPY:
+        # KL divergence
+        kl_div = stats.entropy(p, q)
+        # Jensen-Shannon divergence (symmetric, bounded [0, 1])
+        m = 0.5 * (p + q)
+        js_div = 0.5 * stats.entropy(p, m) + 0.5 * stats.entropy(q, m)
+    else:
+        # Fallback implementations
+        kl_div = float(np.sum(p * np.log(p / q)))
+        m = 0.5 * (p + q)
+        js_div = 0.5 * np.sum(p * np.log(p / m)) + 0.5 * np.sum(q * np.log(q / m))
+
+    # Total variation distance
+    tv_dist = 0.5 * np.sum(np.abs(p - q))
+
+    # Combined drift (JS divergence is bounded [0, ln(2)])
+    drift_score = js_div / np.log(2)  # Normalize to [0, 1]
+
+    return VerificationScore(
+        drift_score=float(np.clip(drift_score, 0.0, 1.0)),
+        confidence=0.9,
+        drift_type=DriftType.NUMERICAL,
+        details={
+            "kl_divergence": float(kl_div),
+            "js_divergence": float(js_div),
+            "total_variation": float(tv_dist),
+        },
+    )
+
+
+def verify_sequences(seq_a: Sequence[str], seq_b: Sequence[str]) -> VerificationScore:
+    """
+    Calculate drift between two sequences of tokens/items.
+
+    Uses edit distance and sequence alignment metrics.
+
+    Args:
+        seq_a: First sequence
+        seq_b: Second sequence
+
+    Returns:
+        VerificationScore with sequence-based drift score
+    """
+    if not seq_a and not seq_b:
+        return VerificationScore(
+            drift_score=0.0,
+            confidence=1.0,
+            drift_type=DriftType.LEXICAL,
+            details={"reason": "both_empty"},
+        )
+
+    # Levenshtein distance
+    edit_dist = _levenshtein_distance(seq_a, seq_b)
+    max_len = max(len(seq_a), len(seq_b))
+    normalized_edit = edit_dist / max_len if max_len > 0 else 0.0
+
+    # Jaccard similarity (set-based)
+    set_a = set(seq_a)
+    set_b = set(seq_b)
+    intersection = len(set_a & set_b)
+    union = len(set_a | set_b)
+    jaccard = intersection / union if union > 0 else 1.0
+    jaccard_drift = 1.0 - jaccard
+
+    # Order-aware similarity (longest common subsequence)
+    lcs_len = _lcs_length(seq_a, seq_b)
+    lcs_ratio = 2 * lcs_len / (len(seq_a) + len(seq_b)) if (len(seq_a) + len(seq_b)) > 0 else 1.0
+    lcs_drift = 1.0 - lcs_ratio
+
+    # Combined
+    drift_score = 0.4 * normalized_edit + 0.3 * jaccard_drift + 0.3 * lcs_drift
+
+    return VerificationScore(
+        drift_score=float(np.clip(drift_score, 0.0, 1.0)),
+        confidence=0.85,
+        drift_type=DriftType.STRUCTURAL,
+        details={
+            "edit_distance": edit_dist,
+            "normalized_edit": float(normalized_edit),
+            "jaccard_similarity": float(jaccard),
+            "lcs_ratio": float(lcs_ratio),
+        },
+    )
+
+
+# ============================================================================
+# Internal pure functions
+# ============================================================================
+
+
+def _lexical_drift(text_a: str, text_b: str) -> dict:
+    """
+    Calculate lexical drift between two texts.
+
+    Pure function - no side effects.
+    """
+    # Character-level comparison
+    chars_a = set(text_a)
+    chars_b = set(text_b)
+    char_jaccard = len(chars_a & chars_b) / len(chars_a | chars_b) if (chars_a | chars_b) else 1.0
+
+    # Word-level comparison
+    words_a = set(text_a.split())
+    words_b = set(text_b.split())
+    word_jaccard = len(words_a & words_b) / len(words_a | words_b) if (words_a | words_b) else 1.0
+
+    # Length ratio
+    len_a, len_b = len(text_a), len(text_b)
+    length_ratio = min(len_a, len_b) / max(len_a, len_b) if max(len_a, len_b) > 0 else 1.0
+
+    # Combined score (lower similarity = higher drift)
+    similarity = 0.3 * char_jaccard + 0.5 * word_jaccard + 0.2 * length_ratio
+    drift = 1.0 - similarity
+
+    return {
+        "score": drift,
+        "char_jaccard": char_jaccard,
+        "word_jaccard": word_jaccard,
+        "length_ratio": length_ratio,
+    }
+
+
+def _structural_drift(text_a: str, text_b: str) -> dict:
+    """
+    Calculate structural drift between two texts.
+
+    Analyzes structure like line count, indentation, code patterns.
+    Pure function - no side effects.
+    """
+    lines_a = text_a.split("\n")
+    lines_b = text_b.split("\n")
+
+    # Line count difference
+    line_count_a, line_count_b = len(lines_a), len(lines_b)
+    line_ratio = (
+        min(line_count_a, line_count_b) / max(line_count_a, line_count_b)
+        if max(line_count_a, line_count_b) > 0
+        else 1.0
+    )
+
+    # Indentation pattern
+    indent_a = [len(line) - len(line.lstrip()) for line in lines_a if line.strip()]
+    indent_b = [len(line) - len(line.lstrip()) for line in lines_b if line.strip()]
+
+    if indent_a and indent_b:
+        avg_indent_a = np.mean(indent_a)
+        avg_indent_b = np.mean(indent_b)
+        max_indent = max(avg_indent_a, avg_indent_b, 1)
+        indent_similarity = 1.0 - abs(avg_indent_a - avg_indent_b) / max_indent
+    else:
+        indent_similarity = 1.0 if (not indent_a and not indent_b) else 0.5
+
+    # Code pattern markers (for code comparison)
+    patterns = ["def ", "class ", "import ", "return ", "if ", "for ", "while ", "try:", "except"]
+    pattern_a = {p for p in patterns if p in text_a}
+    pattern_b = {p for p in patterns if p in text_b}
+    pattern_jaccard = (
+        len(pattern_a & pattern_b) / len(pattern_a | pattern_b) if (pattern_a | pattern_b) else 1.0
+    )
+
+    # Combined
+    similarity = 0.3 * line_ratio + 0.3 * indent_similarity + 0.4 * pattern_jaccard
+    drift = 1.0 - similarity
+
+    return {
+        "score": drift,
+        "line_ratio": line_ratio,
+        "indent_similarity": indent_similarity,
+        "pattern_jaccard": pattern_jaccard,
+    }
+
+
+def _numerical_drift(text_a: str, text_b: str) -> dict:
+    """
+    Calculate numerical drift by extracting and comparing numbers.
+
+    Pure function - no side effects.
+    """
+    import re
+
+    # Extract numbers from both texts
+    number_pattern = r"-?\d+\.?\d*"
+    numbers_a = [float(n) for n in re.findall(number_pattern, text_a)]
+    numbers_b = [float(n) for n in re.findall(number_pattern, text_b)]
+
+    if not numbers_a and not numbers_b:
+        return {"score": 0.0, "reason": "no_numbers"}
+
+    if not numbers_a or not numbers_b:
+        return {"score": 0.5, "reason": "numbers_only_in_one"}
+
+    # Compare statistics
+    mean_a, mean_b = np.mean(numbers_a), np.mean(numbers_b)
+    std_a, std_b = np.std(numbers_a), np.std(numbers_b)
+
+    # Relative difference in means
+    max_mean = max(abs(mean_a), abs(mean_b), 1e-10)
+    mean_diff = abs(mean_a - mean_b) / max_mean
+
+    # Relative difference in stds
+    max_std = max(std_a, std_b, 1e-10)
+    std_diff = abs(std_a - std_b) / max_std if max_std > 1e-10 else 0.0
+
+    # Count difference
+    count_ratio = min(len(numbers_a), len(numbers_b)) / max(len(numbers_a), len(numbers_b))
+
+    # Combined
+    drift = 0.4 * min(mean_diff, 1.0) + 0.3 * min(std_diff, 1.0) + 0.3 * (1.0 - count_ratio)
+
+    return {
+        "score": drift,
+        "mean_a": mean_a,
+        "mean_b": mean_b,
+        "std_a": std_a,
+        "std_b": std_b,
+        "count_a": len(numbers_a),
+        "count_b": len(numbers_b),
+    }
+
+
+def _levenshtein_distance(seq_a: Sequence, seq_b: Sequence) -> int:
+    """
+    Calculate Levenshtein edit distance between two sequences.
+
+    Pure function using dynamic programming.
+    """
+    m, n = len(seq_a), len(seq_b)
+
+    if m == 0:
+        return n
+    if n == 0:
+        return m
+
+    # Use numpy for efficiency
+    dp = np.zeros((m + 1, n + 1), dtype=np.int32)
+    dp[:, 0] = np.arange(m + 1)
+    dp[0, :] = np.arange(n + 1)
+
+    for i in range(1, m + 1):
+        for j in range(1, n + 1):
+            cost = 0 if seq_a[i - 1] == seq_b[j - 1] else 1
+            dp[i, j] = min(
+                dp[i - 1, j] + 1,  # deletion
+                dp[i, j - 1] + 1,  # insertion
+                dp[i - 1, j - 1] + cost,  # substitution
+            )
+
+    return int(dp[m, n])
+
+
+def _lcs_length(seq_a: Sequence, seq_b: Sequence) -> int:
+    """
+    Calculate length of Longest Common Subsequence.
+
+    Pure function using dynamic programming.
+    """
+    m, n = len(seq_a), len(seq_b)
+
+    if m == 0 or n == 0:
+        return 0
+
+    dp = np.zeros((m + 1, n + 1), dtype=np.int32)
+
+    for i in range(1, m + 1):
+        for j in range(1, n + 1):
+            if seq_a[i - 1] == seq_b[j - 1]:
+                dp[i, j] = dp[i - 1, j - 1] + 1
+            else:
+                dp[i, j] = max(dp[i - 1, j], dp[i, j - 1])
+
+    return int(dp[m, n])
+
+
+# ============================================================================
+# Batch verification functions
+# ============================================================================
+
+
+def verify_batch(outputs_a: Sequence[str], outputs_b: Sequence[str]) -> list[VerificationScore]:
+    """
+    Verify multiple output pairs.
+
+    Pure function that processes pairs in sequence.
+
+    Args:
+        outputs_a: Sequence of outputs from source A
+        outputs_b: Sequence of outputs from source B (same length as outputs_a)
+
+    Returns:
+        List of VerificationScore for each pair
+    """
+    if len(outputs_a) != len(outputs_b):
+        raise ValueError(
+            f"Length mismatch: outputs_a has {len(outputs_a)} items, "
+            f"outputs_b has {len(outputs_b)} items"
+        )
+
+    return [verify(a, b) for a, b in zip(outputs_a, outputs_b, strict=False)]
+
+
+def aggregate_scores(scores: Sequence[VerificationScore]) -> dict:
+    """
+    Aggregate multiple verification scores into summary statistics.
+
+    Pure function.
+
+    Args:
+        scores: Sequence of VerificationScore objects
+
+    Returns:
+        Dictionary with aggregate statistics
+    """
+    if not scores:
+        return {"count": 0}
+
+    drift_values = [s.drift_score for s in scores]
+    confidence_values = [s.confidence for s in scores]
+
+    drift_types: dict[str, int] = {}
+    for s in scores:
+        drift_types[s.drift_type.value] = drift_types.get(s.drift_type.value, 0) + 1
+
+    return {
+        "count": len(scores),
+        "mean_drift": float(np.mean(drift_values)),
+        "std_drift": float(np.std(drift_values)),
+        "min_drift": float(np.min(drift_values)),
+        "max_drift": float(np.max(drift_values)),
+        "median_drift": float(np.median(drift_values)),
+        "mean_confidence": float(np.mean(confidence_values)),
+        "drift_type_distribution": drift_types,
+    }