wisent 0.7.379__py3-none-any.whl → 0.7.701__py3-none-any.whl

wisent/core/contrastive_pairs/diagnostics/control_vectors.py

@@ -24,6 +24,17 @@ __all__ = [
     "GeometryAnalysisResult",
     "StructureType",
     "detect_geometry_structure",
+    "MultiLayerGeometryConfig",
+    "MultiLayerGeometryResult",
+    "LayerGeometryResult",
+    "detect_geometry_multi_layer",
+    "detect_geometry_all_layers",
+    "ExhaustiveCombinationResult",
+    "ExhaustiveGeometryAnalysisResult",
+    "detect_geometry_exhaustive",
+    "detect_geometry_limited",
+    "detect_geometry_contiguous",
+    "detect_geometry_smart",
 ]
 
 
@@ -1058,66 +1069,88 @@ def _detect_cone_structure_score(
     neg_tensor: torch.Tensor,
     cfg: GeometryAnalysisConfig,
 ) -> StructureScore:
-    """Detect cone structure and return as StructureScore."""
-    cone_config = ConeAnalysisConfig(
-        num_directions=cfg.num_components,
-        optimization_steps=cfg.optimization_steps,
-        cone_threshold=cfg.cone_threshold,
-    )
+    """Detect cone structure using RAW cosine similarity of difference vectors.
+    
+    A cone structure means:
+    - Multiple difference vectors (pos_i - neg_i) point in SIMILAR directions
+    - High cosine similarity between raw difference vectors
+    - NOT using gradient-optimized directions (which inflate the score)
     
+    This matches what the visualization computes.
+    """
     try:
-        result = check_cone_structure(pos_tensor, neg_tensor, cone_config)
-        
-        # Cone is meaningful when:
-        # 1. Multiple directions are needed (PCA doesn't capture everything)
-        # 2. But directions are correlated (same half-space)
-        # 3. Cosine similarity is moderate (0.3-0.7 range ideal)
-        
-        # Penalize if PCA already explains most variance (that's linear, not cone)
-        pca_penalty = result.pca_explained_variance  # High PCA = linear is enough
-        
-        # Reward if cone explains more than PCA
-        cone_advantage = max(0, result.cone_explained_variance - result.pca_explained_variance)
-        
-        # Cone needs moderate cosine similarity - not too high (= linear) not too low (= orthogonal)
-        cos_sim = result.avg_cosine_similarity
-        if cos_sim > 0.85:
-            # Very high similarity means directions are basically the same = linear
-            cosine_score = 0.3
-        elif cos_sim > 0.7:
-            cosine_score = 0.7
-        elif cos_sim > 0.3:
-            # Ideal range for cone
-            cosine_score = 1.0
+        # Compute raw difference vectors (what visualization uses)
+        n_pairs = min(pos_tensor.shape[0], neg_tensor.shape[0])
+        if n_pairs < 3:
+            return StructureScore(StructureType.CONE, 0.0, 0.0, {"reason": "insufficient_pairs"})
+        
+        diff_vectors = pos_tensor[:n_pairs] - neg_tensor[:n_pairs]
+        
+        # Normalize difference vectors
+        norms = diff_vectors.norm(dim=1, keepdim=True)
+        valid_mask = (norms.squeeze() > 1e-8)
+        if valid_mask.sum() < 3:
+            return StructureScore(StructureType.CONE, 0.0, 0.0, {"reason": "zero_differences"})
+        
+        diff_normalized = diff_vectors[valid_mask] / norms[valid_mask]
+        
+        # Compute pairwise cosine similarity matrix
+        cos_sim_matrix = diff_normalized @ diff_normalized.T
+        
+        # Get off-diagonal elements (exclude self-similarity of 1.0)
+        n = cos_sim_matrix.shape[0]
+        mask = ~torch.eye(n, dtype=torch.bool, device=cos_sim_matrix.device)
+        off_diagonal = cos_sim_matrix[mask]
+        
+        # Raw cosine similarity statistics
+        mean_cos_sim = float(off_diagonal.mean())
+        std_cos_sim = float(off_diagonal.std())
+        min_cos_sim = float(off_diagonal.min())
+        max_cos_sim = float(off_diagonal.max())
+        
+        # Fraction of pairs with positive correlation (same half-space)
+        positive_fraction = float((off_diagonal > 0).float().mean())
+        
+        # Fraction with strong correlation (>0.3)
+        strong_fraction = float((off_diagonal > 0.3).float().mean())
+        
+        # Cone score based on raw cosine similarity:
+        # - High mean cosine = directions are aligned = cone
+        # - Low mean cosine = directions are independent = NOT cone
+        # - Negative mean cosine = directions are opposing = NOT cone
+        
+        if mean_cos_sim < 0:
+            # Negative correlation = definitely not a cone
+            cone_score = 0.0
+        elif mean_cos_sim < 0.1:
+            # Near zero = orthogonal/independent, not cone
+            cone_score = mean_cos_sim  # 0.0 - 0.1
+        elif mean_cos_sim < 0.3:
+            # Weak correlation = weak cone
+            cone_score = 0.1 + 0.2 * ((mean_cos_sim - 0.1) / 0.2)  # 0.1 - 0.3
+        elif mean_cos_sim < 0.7:
+            # Moderate correlation = good cone (ideal range)
+            cone_score = 0.3 + 0.5 * ((mean_cos_sim - 0.3) / 0.4)  # 0.3 - 0.8
         else:
-            # Too different = not a cone
-            cosine_score = max(0, cos_sim / 0.3)
-        
-        # Multiple significant directions needed
-        significant_dirs = sum(1 for s in result.separation_scores if abs(s) > 0.1)
-        multi_dir_score = min(significant_dirs / cfg.num_components, 1.0)
-        
-        # Adjusted cone score
-        cone_score = (
-            0.25 * result.half_space_consistency +
-            0.25 * cosine_score +
-            0.20 * cone_advantage +
-            0.15 * multi_dir_score +
-            0.15 * (1 - pca_penalty)  # Penalize when PCA is sufficient
-        )
+            # Very high correlation = almost linear, still cone-like
+            cone_score = 0.8 + 0.2 * ((mean_cos_sim - 0.7) / 0.3)  # 0.8 - 1.0
+        
+        # Confidence based on consistency (low std = more consistent = higher confidence)
+        consistency = max(0, 1 - std_cos_sim)
+        confidence = consistency * min(1.0, n_pairs / 20)
         
         return StructureScore(
             StructureType.CONE,
             score=float(cone_score),
-            confidence=result.half_space_consistency,
+            confidence=float(confidence),
             details={
-                "pca_explained": result.pca_explained_variance,
-                "cone_explained": result.cone_explained_variance,
-                "cone_advantage": float(cone_advantage),
-                "avg_cosine_similarity": result.avg_cosine_similarity,
-                "half_space_consistency": result.half_space_consistency,
-                "num_directions": result.num_directions_found,
-                "significant_directions": significant_dirs,
+                "raw_mean_cosine_similarity": mean_cos_sim,
+                "raw_std_cosine_similarity": std_cos_sim,
+                "raw_min_cosine_similarity": min_cos_sim,
+                "raw_max_cosine_similarity": max_cos_sim,
+                "positive_correlation_fraction": positive_fraction,
+                "strong_correlation_fraction": strong_fraction,
+                "n_valid_pairs": int(valid_mask.sum()),
             }
         )
     except Exception as e:
@@ -1130,7 +1163,17 @@ def _detect_cluster_structure(
     diff_vectors: torch.Tensor,
     cfg: GeometryAnalysisConfig,
 ) -> StructureScore:
-    """Detect if activations form discrete clusters."""
+    """Detect if activations form discrete clusters.
+    
+    Cluster structure means:
+    - Data forms DISCRETE, SEPARATED groups
+    - Not just "pos vs neg" (that's trivially 2 clusters)
+    - Actual subgroups within the data
+    
+    Key insight: k-means will ALWAYS find clusters.
+    We need high silhouette AND clear separation to claim clusters.
+    Also, if pos/neg perfectly separate, that's "linear", not "cluster".
+    """
     all_activations = torch.cat([pos_tensor, neg_tensor], dim=0)
     n_samples = all_activations.shape[0]
     
@@ -1143,7 +1186,6 @@ def _detect_cluster_structure(
     
     for k in range(2, min(cfg.max_clusters + 1, n_samples // 2)):
         try:
-            # Simple k-means implementation
             labels, centroids, silhouette = _kmeans_with_silhouette(all_activations, k, max_iters=50)
             silhouette_scores[k] = silhouette
             
@@ -1156,31 +1198,55 @@ def _detect_cluster_structure(
     if best_silhouette < 0:
         return StructureScore(StructureType.CLUSTER, 0.0, 0.0, {"reason": "clustering_failed"})
     
-    # Check if clusters separate pos/neg
+    # Check if clusters just separate pos/neg (that's linear, not cluster)
     labels, _, _ = _kmeans_with_silhouette(all_activations, best_k, max_iters=50)
     pos_labels = labels[:pos_tensor.shape[0]]
     neg_labels = labels[pos_tensor.shape[0]:]
     
-    # Cluster purity: do pos and neg end up in different clusters?
-    pos_majority = pos_labels.mode().values.item() if len(pos_labels) > 0 else -1
-    neg_majority = neg_labels.mode().values.item() if len(neg_labels) > 0 else -1
-    cluster_separation = 1.0 if pos_majority != neg_majority else 0.5
-    
-    # Silhouette score ranges from -1 to 1, where:
-    # > 0.7 = strong structure
-    # 0.5-0.7 = reasonable structure
-    # 0.25-0.5 = weak structure
-    # < 0.25 = no substantial structure
-    
-    # Only consider cluster structure if silhouette is reasonably high
-    if best_silhouette < cfg.cluster_silhouette_threshold:
-        # Low silhouette means no clear cluster structure
-        cluster_score = best_silhouette * 0.5  # Scale down significantly
+    # If k=2 and it perfectly separates pos/neg, that's LINEAR not cluster
+    if best_k == 2:
+        pos_mode = pos_labels.mode().values.item() if len(pos_labels) > 0 else -1
+        neg_mode = neg_labels.mode().values.item() if len(neg_labels) > 0 else -1
+        pos_purity = (pos_labels == pos_mode).float().mean()
+        neg_purity = (neg_labels == neg_mode).float().mean()
+        
+        if pos_mode != neg_mode and pos_purity > 0.8 and neg_purity > 0.8:
+            # Perfect pos/neg separation - this is LINEAR, not cluster
+            return StructureScore(
+                StructureType.CLUSTER,
+                score=0.1,  # Low score - it's actually linear
+                confidence=0.8,
+                details={
+                    "reason": "pos_neg_separation_is_linear",
+                    "best_k": 2,
+                    "pos_purity": float(pos_purity),
+                    "neg_purity": float(neg_purity),
+                }
+            )
+    
+    # For true cluster structure, we need:
+    # 1. High silhouette (> 0.5 is good, > 0.7 is strong)
+    # 2. k > 2 OR k=2 with mixed clusters
+    
+    # Silhouette thresholds - be strict
+    if best_silhouette < 0.4:
+        # Very low silhouette = no clear cluster structure
+        cluster_score = best_silhouette * 0.3  # Very low score
+    elif best_silhouette < cfg.cluster_silhouette_threshold:
+        # Moderate silhouette = weak cluster structure
+        cluster_score = 0.1 + 0.2 * (best_silhouette / cfg.cluster_silhouette_threshold)
     else:
-        # Good silhouette - this is truly clustered data
-        # Normalize silhouette from [threshold, 1] to [0.5, 1]
-        normalized_silhouette = (best_silhouette - cfg.cluster_silhouette_threshold) / (1 - cfg.cluster_silhouette_threshold)
-        cluster_score = 0.5 + 0.4 * normalized_silhouette + 0.1 * cluster_separation
+        # High silhouette = good cluster structure
+        # But only if it's not just pos/neg separation
+        base_score = 0.3 + 0.5 * ((best_silhouette - cfg.cluster_silhouette_threshold) / (1 - cfg.cluster_silhouette_threshold))
+        
+        # Bonus for k > 2 (more interesting structure)
+        if best_k > 2:
+            cluster_score = base_score + 0.2
+        else:
+            cluster_score = base_score
+    
+    cluster_score = min(1.0, cluster_score)
     
     return StructureScore(
         StructureType.CLUSTER,
@@ -1190,7 +1256,6 @@ def _detect_cluster_structure(
             "best_k": best_k,
             "best_silhouette": float(best_silhouette),
             "all_silhouettes": {str(k): float(v) for k, v in silhouette_scores.items()},
-            "cluster_separation": float(cluster_separation),
             "silhouette_threshold": cfg.cluster_silhouette_threshold,
         }
     )
@@ -1279,7 +1344,19 @@ def _detect_manifold_structure(
     diff_vectors: torch.Tensor,
     cfg: GeometryAnalysisConfig,
 ) -> StructureScore:
-    """Detect non-linear manifold structure via intrinsic dimensionality."""
+    """Detect non-linear manifold structure.
+    
+    Manifold structure means:
+    - Data lies on a CURVED surface (not linear)
+    - Linear methods (PCA, CAA) cannot capture the structure
+    - Requires non-linear methods (TITAN, neural steering)
+    
+    Key insight: Manifold should be a FALLBACK, not default.
+    Only report manifold if:
+    1. Linear doesn't work (PCA explains little variance)
+    2. There's actual curvature (local neighborhoods don't align)
+    3. BUT there IS structure (not just noise)
+    """
     all_activations = torch.cat([pos_tensor, neg_tensor], dim=0)
     n_samples = all_activations.shape[0]
     
@@ -1287,52 +1364,80 @@ def _detect_manifold_structure(
         return StructureScore(StructureType.MANIFOLD, 0.0, 0.0, {"reason": "insufficient_data"})
     
     try:
-        # First check if there's meaningful separation
+        # 1. Check if linear works well (if yes, not manifold)
+        centered = all_activations - all_activations.mean(dim=0, keepdim=True)
+        try:
+            _, S, _ = torch.linalg.svd(centered, full_matrices=False)
+            total_var = (S ** 2).sum()
+            if total_var > 0:
+                # Top 2 PCs variance explained
+                top2_var = (S[:2] ** 2).sum() / total_var
+                linear_explains_well = float(top2_var) > 0.7
+            else:
+                linear_explains_well = True  # No variance = trivial
+        except Exception:
+            linear_explains_well = False
+            top2_var = torch.tensor(0.0)
+        
+        if linear_explains_well:
+            # Linear works well - not a manifold (it's linear)
+            return StructureScore(
+                StructureType.MANIFOLD, 
+                score=0.1, 
+                confidence=0.8,
+                details={
+                    "reason": "linear_sufficient",
+                    "pca_top2_variance": float(top2_var),
+                }
+            )
+        
+        # 2. Check for actual curvature (local PCA directions vary)
+        local_nonlinearity = _compute_local_nonlinearity(all_activations, cfg.manifold_neighbors)
+        
+        # 3. Check if there's meaningful structure (separation between pos/neg)
         mean_diff = pos_tensor.mean(dim=0) - neg_tensor.mean(dim=0)
         separation_strength = mean_diff.norm() / (pos_tensor.std() + neg_tensor.std() + 1e-8)
         has_structure = min(float(separation_strength) / 2, 1.0)
         
-        if has_structure < 0.2:
-            # No meaningful separation - can't determine manifold structure
-            return StructureScore(StructureType.MANIFOLD, 0.1, 0.0, {"reason": "no_separation"})
-        
-        # Estimate intrinsic dimensionality using correlation dimension
-        intrinsic_dim = _estimate_intrinsic_dimensionality(all_activations, cfg.manifold_neighbors)
-        
-        # Compare to ambient dimension
-        ambient_dim = all_activations.shape[1]
-        dim_ratio = intrinsic_dim / ambient_dim
-        
-        # Also compute local linearity deviation
-        local_nonlinearity = _compute_local_nonlinearity(all_activations, cfg.manifold_neighbors)
+        if has_structure < 0.3:
+            # No clear structure - likely noise, not manifold
+            return StructureScore(
+                StructureType.MANIFOLD,
+                score=0.2,
+                confidence=0.5,
+                details={
+                    "reason": "weak_structure",
+                    "separation_strength": float(separation_strength),
+                }
+            )
         
-        # Manifold score: high if low intrinsic dim AND non-linear AND has structure
-        # Low intrinsic dim alone could be linear, so we need nonlinearity
-        # But random noise also has "nonlinearity" - need to distinguish
+        # 4. Manifold requires BOTH:
+        #    - Linear doesn't work (already checked)
+        #    - AND there's curvature
+        #    - AND there's structure
         
-        # Manifold is meaningful only with significant dimension reduction
-        if dim_ratio > 0.5:
-            # Not much dimension reduction = not a clear manifold
-            manifold_score = 0.3 * has_structure
+        # If nonlinearity is low, it might be orthogonal/independent, not curved
+        if local_nonlinearity < 0.3:
+            manifold_score = 0.3 * has_structure  # Low score
         else:
+            # High nonlinearity + structure = manifold candidate
             manifold_score = (
-                0.30 * (1 - dim_ratio) +
-                0.25 * local_nonlinearity +
-                0.45 * has_structure  # Weight structure heavily
+                0.30 * local_nonlinearity +
+                0.30 * (1 - float(top2_var)) +  # Reward when linear fails
+                0.40 * has_structure
             )
         
-        # Confidence based on sample size
-        confidence = min(1.0, n_samples / 100)
+        # Confidence based on sample size and consistency
+        confidence = min(1.0, n_samples / 100) * has_structure
         
         return StructureScore(
             StructureType.MANIFOLD,
             score=float(manifold_score),
             confidence=float(confidence),
             details={
-                "intrinsic_dimensionality": float(intrinsic_dim),
-                "ambient_dimensionality": ambient_dim,
-                "dim_ratio": float(dim_ratio),
+                "pca_top2_variance": float(top2_var),
                 "local_nonlinearity": float(local_nonlinearity),
+                "separation_strength": float(separation_strength),
             }
         )
     except Exception as e:
@@ -1547,83 +1652,85 @@ def _detect_orthogonal_structure(
 ) -> StructureScore:
     """Detect if behavior is encoded in multiple orthogonal/independent subspaces.
     
-    Orthogonal structure means the data requires MULTIPLE independent directions
-    that are NOT correlated with each other. This is different from cone (where
-    directions are correlated) and linear (where one direction suffices).
-    """
-    if diff_vectors.shape[0] < cfg.num_components:
-        return StructureScore(StructureType.ORTHOGONAL, 0.0, 0.0, {"reason": "insufficient_data"})
+    Orthogonal structure means:
+    - Multiple difference vectors point in INDEPENDENT directions
+    - Low cosine similarity between difference vectors (near 0)
+    - NOT correlated (that's cone) and NOT single direction (that's linear)
     
+    This is the OPPOSITE of cone - if cosine sim is low, it's orthogonal.
+    Uses raw cosine similarity like the cone detector for consistency.
+    """
     try:
-        # PCA to understand variance distribution
-        centered = diff_vectors - diff_vectors.mean(dim=0, keepdim=True)
-        _, S, Vh = torch.linalg.svd(centered, full_matrices=False)
-        
-        total_var = (S ** 2).sum()
-        if total_var < 1e-8:
-            return StructureScore(StructureType.ORTHOGONAL, 0.0, 0.0, {"reason": "no_variance"})
-        
-        # For orthogonal structure:
-        # 1. Multiple components should have significant variance (not just one = linear)
-        # 2. Variance should be spread across multiple dimensions (not concentrated)
-        
-        var_explained = (S ** 2) / total_var
-        k = min(cfg.num_components, len(S))
-        
-        # First component dominance (low = more orthogonal/spread)
-        first_var = float(var_explained[0])
-        
-        # Effective dimensionality (entropy-based)
-        var_explained_clipped = var_explained[var_explained > 1e-10]
-        entropy = -(var_explained_clipped * torch.log(var_explained_clipped + 1e-10)).sum()
-        max_entropy = torch.log(torch.tensor(float(len(var_explained_clipped))))
-        effective_dim_ratio = float(entropy / max_entropy) if max_entropy > 0 else 0.0
-        
-        # Count significant dimensions (>5% variance each)
-        significant_dims = (var_explained > 0.05).sum().item()
-        multi_dim_score = min(significant_dims / 3, 1.0)  # 3+ significant dims is fully orthogonal
-        
-        # Orthogonal structure is RARE and specific:
-        # It requires MULTIPLE INDEPENDENT directions with separation on EACH
-        # High spread alone is not orthogonal - it could be noise or cone
+        # Compute raw difference vectors (same as cone detector)
+        n_pairs = min(pos_tensor.shape[0], neg_tensor.shape[0])
+        if n_pairs < 3:
+            return StructureScore(StructureType.ORTHOGONAL, 0.0, 0.0, {"reason": "insufficient_pairs"})
+        
+        diff_vectors_raw = pos_tensor[:n_pairs] - neg_tensor[:n_pairs]
+        
+        # Normalize difference vectors
+        norms = diff_vectors_raw.norm(dim=1, keepdim=True)
+        valid_mask = (norms.squeeze() > 1e-8)
+        if valid_mask.sum() < 3:
+            return StructureScore(StructureType.ORTHOGONAL, 0.0, 0.0, {"reason": "zero_differences"})
+        
+        diff_normalized = diff_vectors_raw[valid_mask] / norms[valid_mask]
+        
+        # Compute pairwise cosine similarity matrix
+        cos_sim_matrix = diff_normalized @ diff_normalized.T
+        
+        # Get off-diagonal elements
+        n = cos_sim_matrix.shape[0]
+        mask = ~torch.eye(n, dtype=torch.bool, device=cos_sim_matrix.device)
+        off_diagonal = cos_sim_matrix[mask]
+        
+        # Raw cosine similarity statistics
+        mean_cos_sim = float(off_diagonal.mean())
+        std_cos_sim = float(off_diagonal.std())
+        abs_mean_cos_sim = float(off_diagonal.abs().mean())
+        
+        # Fraction near zero (truly orthogonal)
+        near_zero_fraction = float((off_diagonal.abs() < 0.2).float().mean())
+        
+        # Orthogonal = LOW cosine similarity (opposite of cone)
+        # Ideal orthogonal: mean cosine sim near 0, low absolute mean
+        
+        if abs_mean_cos_sim < 0.1:
+            # Very low correlation = strong orthogonal
+            orthogonal_score = 0.8 + 0.2 * (1 - abs_mean_cos_sim / 0.1)
+        elif abs_mean_cos_sim < 0.2:
+            # Low correlation = moderate orthogonal
+            orthogonal_score = 0.5 + 0.3 * (1 - (abs_mean_cos_sim - 0.1) / 0.1)
+        elif abs_mean_cos_sim < 0.4:
+            # Moderate correlation = weak orthogonal
+            orthogonal_score = 0.2 + 0.3 * (1 - (abs_mean_cos_sim - 0.2) / 0.2)
+        else:
+            # High correlation = not orthogonal (probably cone or linear)
+            orthogonal_score = max(0, 0.2 * (1 - (abs_mean_cos_sim - 0.4) / 0.6))
         
-        # Check separation strength
+        # Check if there's meaningful separation (not just noise)
         mean_diff = pos_tensor.mean(dim=0) - neg_tensor.mean(dim=0)
         separation_strength = mean_diff.norm() / (pos_tensor.std() + neg_tensor.std() + 1e-8)
-        has_separation = min(float(separation_strength) / 3, 1.0)
-        
-        # For true orthogonal structure, we need:
-        # 1. Strong separation (otherwise no structure)
-        # 2. Multiple significant dimensions (otherwise linear)
-        # 3. But NOT too spread (otherwise just noise)
-        
-        # Sweet spot: 2-4 significant dimensions with clear separation
-        if significant_dims < 2:
-            # Too few dimensions = linear
-            orthogonal_score = 0.2
-        elif significant_dims > 10:
-            # Too many = likely noise, not structure
-            orthogonal_score = 0.3 * has_separation
-        else:
-            # Reasonable number of dimensions
-            # Check if it's not dominated by first (would be linear)
-            # and not too spread (would be noise)
-            structure_score = (
-                0.3 * (1 - first_var) +  # Not dominated by one direction
-                0.3 * min(significant_dims / 4, 1.0) +  # 2-4 directions is ideal
-                0.4 * has_separation  # Must have separation
-            )
-            orthogonal_score = structure_score * 0.8  # Scale down - orthogonal is rare
+        has_separation = min(float(separation_strength) / 2, 1.0)
+        
+        # Orthogonal without separation is just noise
+        if has_separation < 0.3:
+            orthogonal_score *= 0.3  # Heavy penalty
+        
+        # Confidence based on consistency and sample size
+        confidence = near_zero_fraction * min(1.0, n_pairs / 20)
         
         return StructureScore(
             StructureType.ORTHOGONAL,
             score=float(orthogonal_score),
-            confidence=min(1.0, diff_vectors.shape[0] / 30),
+            confidence=float(confidence),
             details={
-                "first_component_variance": float(first_var),
-                "effective_dim_ratio": float(effective_dim_ratio),
-                "significant_dimensions": int(significant_dims),
-                "top_5_variances": var_explained[:min(5, len(var_explained))].tolist(),
+                "raw_mean_cosine_similarity": mean_cos_sim,
+                "raw_abs_mean_cosine_similarity": abs_mean_cos_sim,
+                "raw_std_cosine_similarity": std_cos_sim,
+                "near_zero_fraction": near_zero_fraction,
+                "separation_strength": float(separation_strength),
+                "n_valid_pairs": int(valid_mask.sum()),
             }
         )
     except Exception as e:
@@ -1652,4 +1759,1302 @@ def _generate_recommendation(best_structure: StructureType, all_scores: Dict[str
         if second_best[1].score > 0.6:
             base_rec += f" (Also consider {second_best[0]}: score {second_best[1].score:.2f})"
     
-    return base_rec
+    return base_rec
+
+
+# =============================================================================
+# Multi-Layer Geometry Analysis
+# =============================================================================
+
+@dataclass
+class MultiLayerGeometryConfig:
+    """Configuration for multi-layer geometry analysis."""
+    
+    num_components: int = 5
+    optimization_steps: int = 50
+    combination_method: str = "concat"  # "concat", "mean", "weighted"
+    analyze_per_layer: bool = True
+    analyze_combined: bool = True
+    analyze_subsets: bool = True  # early/middle/late
+    analyze_pairs: bool = True  # all pairs of layers
+    analyze_adjacent: bool = True  # adjacent layer pairs
+    analyze_skip: bool = True  # every other layer, every third, etc.
+    analyze_custom: Optional[List[List[int]]] = None  # custom layer combinations
+    max_pair_combinations: int = 50  # limit number of pair combinations to analyze
+
+
+@dataclass
+class LayerGeometryResult:
+    """Geometry result for a single layer."""
+    layer: int
+    best_structure: StructureType
+    best_score: float
+    all_scores: Dict[str, float]
+
+
+@dataclass
+class MultiLayerGeometryResult:
+    """Results from multi-layer geometry analysis."""
+    
+    per_layer_results: Dict[int, LayerGeometryResult]
+    """Geometry analysis for each individual layer."""
+    
+    combined_result: Optional[GeometryAnalysisResult]
+    """Geometry analysis for all layers combined."""
+    
+    layer_subset_results: Dict[str, GeometryAnalysisResult]
+    """Geometry analysis for layer subsets (e.g., 'early', 'middle', 'late')."""
+    
+    layer_pair_results: Dict[str, GeometryAnalysisResult]
+    """Geometry analysis for pairs of layers (e.g., 'L1+L5', 'L2+L8')."""
+    
+    adjacent_pair_results: Dict[str, GeometryAnalysisResult]
+    """Geometry analysis for adjacent layer pairs (e.g., 'L1+L2', 'L2+L3')."""
+    
+    skip_results: Dict[str, GeometryAnalysisResult]
+    """Geometry analysis for skip patterns (e.g., 'every_2nd', 'every_3rd')."""
+    
+    custom_results: Dict[str, GeometryAnalysisResult]
+    """Geometry analysis for custom layer combinations."""
+    
+    best_single_layer: int
+    """Layer with strongest structure detection."""
+    
+    best_single_layer_structure: StructureType
+    """Structure type detected at best single layer."""
+    
+    best_single_layer_score: float
+    """Score at best single layer."""
+    
+    best_combination: Optional[str]
+    """Best performing layer combination (if better than single layer)."""
+    
+    best_combination_score: float
+    """Score of best combination."""
+    
+    best_combination_structure: Optional[StructureType]
+    """Structure type detected at best combination."""
+    
+    combined_vs_single: str
+    """Whether combined layers improve over single layer."""
+    
+    layer_agreement: float
+    """How much layers agree on structure type (0-1)."""
+    
+    structure_by_depth: Dict[str, List[float]]
+    """How each structure score varies by layer depth."""
+    
+    all_combinations_ranked: List[Tuple[str, float, StructureType]]
+    """All combinations ranked by score: (name, score, structure)."""
+    
+    recommendation: str
+    """Recommendation based on multi-layer analysis."""
+
+
+def detect_geometry_multi_layer(
+    pos_activations_by_layer: Dict[int, torch.Tensor],
+    neg_activations_by_layer: Dict[int, torch.Tensor],
+    config: MultiLayerGeometryConfig | None = None,
+) -> MultiLayerGeometryResult:
+    """
+    Detect geometric structure across multiple layers.
+    
+    Analyzes:
+    1. Each layer individually
+    2. All layers combined (concatenated or aggregated)
+    3. Layer subsets (early, middle, late)
+    4. Layer pairs (all combinations of 2 layers)
+    5. Adjacent layer pairs (L1+L2, L2+L3, etc.)
+    6. Skip patterns (every 2nd, every 3rd layer)
+    7. Custom layer combinations
+    8. How structure varies by depth
+    
+    Arguments:
+        pos_activations_by_layer: Dict mapping layer index to positive activations [N, hidden_dim]
+        neg_activations_by_layer: Dict mapping layer index to negative activations [N, hidden_dim]
+        config: Analysis configuration
+        
+    Returns:
+        MultiLayerGeometryResult with comprehensive multi-layer analysis
+    """
+    cfg = config or MultiLayerGeometryConfig()
+    geo_cfg = GeometryAnalysisConfig(num_components=cfg.num_components, optimization_steps=cfg.optimization_steps)
+    
+    layers = sorted(pos_activations_by_layer.keys())
+    if not layers:
+        raise ValueError("No layers provided")
+    
+    # Track all combination results for ranking
+    all_combo_results: Dict[str, GeometryAnalysisResult] = {}
+    
+    # 1. Analyze each layer individually
+    per_layer_results: Dict[int, LayerGeometryResult] = {}
+    structure_by_depth: Dict[str, List[float]] = {
+        "linear": [], "cone": [], "cluster": [], "manifold": [],
+        "sparse": [], "bimodal": [], "orthogonal": []
+    }
+    
+    if cfg.analyze_per_layer:
+        for layer in layers:
+            pos_acts = pos_activations_by_layer[layer]
+            neg_acts = neg_activations_by_layer[layer]
+            
+            result = detect_geometry_structure(pos_acts, neg_acts, geo_cfg)
+            
+            all_scores = {name: score.score for name, score in result.all_scores.items()}
+            per_layer_results[layer] = LayerGeometryResult(
+                layer=layer,
+                best_structure=result.best_structure,
+                best_score=result.best_score,
+                all_scores=all_scores,
+            )
+            all_combo_results[f"L{layer}"] = result
+            
+            for struct_name, score in all_scores.items():
+                if struct_name in structure_by_depth:
+                    structure_by_depth[struct_name].append(score)
+    
+    # 2. Find best single layer
+    if per_layer_results:
+        best_layer = max(per_layer_results.keys(), key=lambda l: per_layer_results[l].best_score)
+        best_single_layer = best_layer
+        best_single_layer_structure = per_layer_results[best_layer].best_structure
+        best_single_layer_score = per_layer_results[best_layer].best_score
+    else:
+        best_single_layer = layers[0]
+        best_single_layer_structure = StructureType.UNKNOWN
+        best_single_layer_score = 0.0
+    
+    # 3. Analyze all layers combined
+    combined_result = None
+    if cfg.analyze_combined and len(layers) > 1:
+        combined_pos, combined_neg = _combine_layer_activations(
+            pos_activations_by_layer, neg_activations_by_layer, layers, cfg.combination_method
+        )
+        combined_result = detect_geometry_structure(combined_pos, combined_neg, geo_cfg)
+        all_combo_results["all_layers"] = combined_result
+    
+    # 4. Analyze layer subsets (early, middle, late)
+    layer_subset_results: Dict[str, GeometryAnalysisResult] = {}
+    if cfg.analyze_subsets and len(layers) >= 3:
+        n_layers = len(layers)
+        third = n_layers // 3
+        
+        early_layers = layers[:third] if third > 0 else layers[:1]
+        middle_layers = layers[third:2*third] if third > 0 else layers[1:2]
+        late_layers = layers[2*third:] if third > 0 else layers[-1:]
+        
+        # Also add first_half and second_half
+        half = n_layers // 2
+        first_half = layers[:half] if half > 0 else layers[:1]
+        second_half = layers[half:] if half > 0 else layers[-1:]
+        
+        subsets = [
+            ("early", early_layers),
+            ("middle", middle_layers),
+            ("late", late_layers),
+            ("first_half", first_half),
+            ("second_half", second_half),
+        ]
+        
+        for subset_name, subset_layers in subsets:
+            if len(subset_layers) >= 1:
+                subset_pos, subset_neg = _combine_layer_activations(
+                    pos_activations_by_layer, neg_activations_by_layer, subset_layers, cfg.combination_method
+                )
+                result = detect_geometry_structure(subset_pos, subset_neg, geo_cfg)
+                layer_subset_results[subset_name] = result
+                all_combo_results[subset_name] = result
+    
+    # 5. Analyze layer pairs
+    layer_pair_results: Dict[str, GeometryAnalysisResult] = {}
+    if cfg.analyze_pairs and len(layers) >= 2:
+        from itertools import combinations
+        pair_count = 0
+        for l1, l2 in combinations(layers, 2):
+            if pair_count >= cfg.max_pair_combinations:
+                break
+            pair_name = f"L{l1}+L{l2}"
+            pair_pos, pair_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, [l1, l2], cfg.combination_method
+            )
+            result = detect_geometry_structure(pair_pos, pair_neg, geo_cfg)
+            layer_pair_results[pair_name] = result
+            all_combo_results[pair_name] = result
+            pair_count += 1
+    
+    # 6. Analyze adjacent layer pairs
+    adjacent_pair_results: Dict[str, GeometryAnalysisResult] = {}
+    if cfg.analyze_adjacent and len(layers) >= 2:
+        for i in range(len(layers) - 1):
+            l1, l2 = layers[i], layers[i + 1]
+            pair_name = f"adj_L{l1}+L{l2}"
+            pair_pos, pair_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, [l1, l2], cfg.combination_method
+            )
+            result = detect_geometry_structure(pair_pos, pair_neg, geo_cfg)
+            adjacent_pair_results[pair_name] = result
+            all_combo_results[pair_name] = result
+    
+    # 7. Analyze skip patterns
+    skip_results: Dict[str, GeometryAnalysisResult] = {}
+    if cfg.analyze_skip and len(layers) >= 4:
+        # Every 2nd layer
+        every_2nd = layers[::2]
+        if len(every_2nd) >= 2:
+            skip_pos, skip_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, every_2nd, cfg.combination_method
+            )
+            result = detect_geometry_structure(skip_pos, skip_neg, geo_cfg)
+            skip_results["every_2nd"] = result
+            all_combo_results["every_2nd"] = result
+        
+        # Every 3rd layer
+        if len(layers) >= 6:
+            every_3rd = layers[::3]
+            if len(every_3rd) >= 2:
+                skip_pos, skip_neg = _combine_layer_activations(
+                    pos_activations_by_layer, neg_activations_by_layer, every_3rd, cfg.combination_method
+                )
+                result = detect_geometry_structure(skip_pos, skip_neg, geo_cfg)
+                skip_results["every_3rd"] = result
+                all_combo_results["every_3rd"] = result
+        
+        # First and last layer only
+        first_last = [layers[0], layers[-1]]
+        skip_pos, skip_neg = _combine_layer_activations(
+            pos_activations_by_layer, neg_activations_by_layer, first_last, cfg.combination_method
+        )
+        result = detect_geometry_structure(skip_pos, skip_neg, geo_cfg)
+        skip_results["first_last"] = result
+        all_combo_results["first_last"] = result
+        
+        # First, middle, last
+        if len(layers) >= 3:
+            mid_idx = len(layers) // 2
+            first_mid_last = [layers[0], layers[mid_idx], layers[-1]]
+            skip_pos, skip_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, first_mid_last, cfg.combination_method
+            )
+            result = detect_geometry_structure(skip_pos, skip_neg, geo_cfg)
+            skip_results["first_mid_last"] = result
+            all_combo_results["first_mid_last"] = result
+    
+    # 8. Analyze custom combinations
+    custom_results: Dict[str, GeometryAnalysisResult] = {}
+    if cfg.analyze_custom:
+        for i, custom_layers in enumerate(cfg.analyze_custom):
+            valid_layers = [l for l in custom_layers if l in layers]
+            if len(valid_layers) >= 1:
+                custom_name = f"custom_{i}_L" + "+L".join(map(str, valid_layers))
+                custom_pos, custom_neg = _combine_layer_activations(
+                    pos_activations_by_layer, neg_activations_by_layer, valid_layers, cfg.combination_method
+                )
+                result = detect_geometry_structure(custom_pos, custom_neg, geo_cfg)
+                custom_results[custom_name] = result
+                all_combo_results[custom_name] = result
+    
+    # 9. Compute layer agreement
+    if per_layer_results:
+        structures = [r.best_structure for r in per_layer_results.values()]
+        most_common = max(set(structures), key=structures.count)
+        layer_agreement = structures.count(most_common) / len(structures)
+    else:
+        layer_agreement = 0.0
+    
+    # 10. Rank all combinations and find best
+    all_combinations_ranked = sorted(
+        [(name, r.best_score, r.best_structure) for name, r in all_combo_results.items()],
+        key=lambda x: x[1],
+        reverse=True
+    )
+    
+    if all_combinations_ranked:
+        best_combo_name, best_combo_score, best_combo_structure = all_combinations_ranked[0]
+        if best_combo_score > best_single_layer_score:
+            best_combination = best_combo_name
+            best_combination_score = best_combo_score
+            best_combination_structure = best_combo_structure
+        else:
+            best_combination = None
+            best_combination_score = best_single_layer_score
+            best_combination_structure = best_single_layer_structure
+    else:
+        best_combination = None
+        best_combination_score = best_single_layer_score
+        best_combination_structure = best_single_layer_structure
+    
+    # 11. Compare combined vs single
+    if combined_result and per_layer_results:
+        if combined_result.best_score > best_single_layer_score + 0.1:
+            combined_vs_single = f"Combined ({combined_result.best_score:.2f}) better than single layer ({best_single_layer_score:.2f})"
+        elif best_single_layer_score > combined_result.best_score + 0.1:
+            combined_vs_single = f"Single layer {best_single_layer} ({best_single_layer_score:.2f}) better than combined ({combined_result.best_score:.2f})"
+        else:
+            combined_vs_single = f"Similar performance: combined={combined_result.best_score:.2f}, single={best_single_layer_score:.2f}"
+    else:
+        combined_vs_single = "No comparison available"
+    
+    # 12. Generate recommendation
+    recommendation = _generate_multi_layer_recommendation_v2(
+        per_layer_results, combined_result, layer_subset_results,
+        layer_pair_results, skip_results,
+        best_single_layer, best_single_layer_structure, best_single_layer_score,
+        best_combination, best_combination_score, best_combination_structure,
+        layer_agreement, all_combinations_ranked
+    )
+    
+    return MultiLayerGeometryResult(
+        per_layer_results=per_layer_results,
+        combined_result=combined_result,
+        layer_subset_results=layer_subset_results,
+        layer_pair_results=layer_pair_results,
+        adjacent_pair_results=adjacent_pair_results,
+        skip_results=skip_results,
+        custom_results=custom_results,
+        best_single_layer=best_single_layer,
+        best_single_layer_structure=best_single_layer_structure,
+        best_single_layer_score=best_single_layer_score,
+        best_combination=best_combination,
+        best_combination_score=best_combination_score,
+        best_combination_structure=best_combination_structure,
+        combined_vs_single=combined_vs_single,
+        layer_agreement=layer_agreement,
+        structure_by_depth=structure_by_depth,
+        all_combinations_ranked=all_combinations_ranked,
+        recommendation=recommendation,
+    )
+
+
+def _combine_layer_activations(
+    pos_by_layer: Dict[int, torch.Tensor],
+    neg_by_layer: Dict[int, torch.Tensor],
+    layers: List[int],
+    method: str = "concat",
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Combine activations from multiple layers."""
+    pos_acts = [pos_by_layer[l] for l in layers if l in pos_by_layer]
+    neg_acts = [neg_by_layer[l] for l in layers if l in neg_by_layer]
+    
+    if not pos_acts or not neg_acts:
+        raise ValueError("No activations found for specified layers")
+    
+    if method == "concat":
+        combined_pos = torch.cat(pos_acts, dim=-1)
+        combined_neg = torch.cat(neg_acts, dim=-1)
+    elif method == "mean":
+        combined_pos = torch.stack(pos_acts, dim=0).mean(dim=0)
+        combined_neg = torch.stack(neg_acts, dim=0).mean(dim=0)
+    elif method == "weighted":
+        weights = torch.linspace(0.5, 1.5, len(pos_acts))
+        weights = weights / weights.sum()
+        combined_pos = sum(w * a for w, a in zip(weights, pos_acts))
+        combined_neg = sum(w * a for w, a in zip(weights, neg_acts))
+    else:
+        raise ValueError(f"Unknown combination method: {method}")
+    
+    return combined_pos, combined_neg
+
+
+def _generate_multi_layer_recommendation(
+    per_layer_results: Dict[int, LayerGeometryResult],
+    combined_result: Optional[GeometryAnalysisResult],
+    layer_subset_results: Dict[str, GeometryAnalysisResult],
+    best_single_layer: int,
+    best_single_layer_structure: StructureType,
+    best_single_layer_score: float,
+    layer_agreement: float,
+) -> str:
+    """Generate recommendation based on multi-layer analysis."""
+    parts = []
+    
+    # Layer agreement insight
+    if layer_agreement > 0.8:
+        parts.append(f"High layer agreement ({layer_agreement:.0%}): structure is consistent across depth.")
+    elif layer_agreement < 0.4:
+        parts.append(f"Low layer agreement ({layer_agreement:.0%}): different structures at different depths.")
+    
+    # Best layer recommendation
+    parts.append(f"Best single layer: {best_single_layer} with {best_single_layer_structure.value} ({best_single_layer_score:.2f}).")
+    
+    # Combined vs single
+    if combined_result:
+        if combined_result.best_score > best_single_layer_score + 0.1:
+            parts.append(f"Combined layers improve detection ({combined_result.best_score:.2f} vs {best_single_layer_score:.2f}). Use multi-layer steering.")
+        else:
+            parts.append(f"Single layer is sufficient. Target layer {best_single_layer}.")
+    
+    # Layer subset insights
+    if layer_subset_results:
+        subset_scores = {name: r.best_score for name, r in layer_subset_results.items()}
+        best_subset = max(subset_scores.keys(), key=lambda k: subset_scores[k])
+        if subset_scores[best_subset] > best_single_layer_score:
+            parts.append(f"'{best_subset}' layers show strongest structure ({subset_scores[best_subset]:.2f}).")
+    
+    return " ".join(parts)
+
+
+def _generate_multi_layer_recommendation_v2(
+    per_layer_results: Dict[int, LayerGeometryResult],
+    combined_result: Optional[GeometryAnalysisResult],
+    layer_subset_results: Dict[str, GeometryAnalysisResult],
+    layer_pair_results: Dict[str, GeometryAnalysisResult],
+    skip_results: Dict[str, GeometryAnalysisResult],
+    best_single_layer: int,
+    best_single_layer_structure: StructureType,
+    best_single_layer_score: float,
+    best_combination: Optional[str],
+    best_combination_score: float,
+    best_combination_structure: Optional[StructureType],
+    layer_agreement: float,
+    all_combinations_ranked: List[Tuple[str, float, StructureType]],
+) -> str:
+    """Generate comprehensive recommendation based on multi-layer analysis."""
+    parts = []
+    
+    # Layer agreement insight
+    if layer_agreement > 0.8:
+        parts.append(f"High layer agreement ({layer_agreement:.0%}): consistent structure across depth.")
+    elif layer_agreement < 0.4:
+        parts.append(f"Low layer agreement ({layer_agreement:.0%}): structure varies by depth.")
+    else:
+        parts.append(f"Moderate layer agreement ({layer_agreement:.0%}).")
+    
+    # Overall best recommendation
+    if best_combination and best_combination_score > best_single_layer_score + 0.05:
+        improvement = best_combination_score - best_single_layer_score
+        parts.append(
+            f"BEST: '{best_combination}' ({best_combination_structure.value}: {best_combination_score:.2f}) "
+            f"outperforms single layer {best_single_layer} by {improvement:.2f}."
+        )
+    else:
+        parts.append(
+            f"BEST: Layer {best_single_layer} ({best_single_layer_structure.value}: {best_single_layer_score:.2f}). "
+            f"Multi-layer combinations don't improve detection."
+        )
+    
+    # Top 3 combinations summary
+    if len(all_combinations_ranked) >= 3:
+        top3 = all_combinations_ranked[:3]
+        top3_str = ", ".join([f"{name}={score:.2f}" for name, score, _ in top3])
+        parts.append(f"Top 3: {top3_str}.")
+    
+    # Specific pattern insights
+    if skip_results:
+        skip_scores = {name: r.best_score for name, r in skip_results.items()}
+        best_skip = max(skip_scores.keys(), key=lambda k: skip_scores[k])
+        if skip_scores[best_skip] > best_single_layer_score:
+            parts.append(f"Skip pattern '{best_skip}' is effective ({skip_scores[best_skip]:.2f}).")
+    
+    if layer_pair_results:
+        pair_scores = {name: r.best_score for name, r in layer_pair_results.items()}
+        best_pair = max(pair_scores.keys(), key=lambda k: pair_scores[k])
+        best_pair_score = pair_scores[best_pair]
+        if best_pair_score > best_single_layer_score:
+            parts.append(f"Layer pair '{best_pair}' shows synergy ({best_pair_score:.2f}).")
+    
+    # Depth pattern analysis
+    if per_layer_results and len(per_layer_results) >= 3:
+        layers_sorted = sorted(per_layer_results.keys())
+        early_score = per_layer_results[layers_sorted[0]].best_score
+        late_score = per_layer_results[layers_sorted[-1]].best_score
+        if late_score > early_score + 0.2:
+            parts.append("Later layers show stronger structure than early layers.")
+        elif early_score > late_score + 0.2:
+            parts.append("Early layers show stronger structure than later layers.")
+    
+    return " ".join(parts)
+
+
+def detect_geometry_all_layers(
+    pairs_with_activations: List,
+    layers: Optional[List[int]] = None,
+    config: MultiLayerGeometryConfig | None = None,
+) -> MultiLayerGeometryResult:
+    """
+    Convenience function to detect geometry from pairs with pre-collected activations.
+    
+    Arguments:
+        pairs_with_activations: List of ContrastivePair objects with layers_activations populated
+        layers: Specific layers to analyze (None = all available)
+        config: Analysis configuration
+        
+    Returns:
+        MultiLayerGeometryResult
+    """
+    if not pairs_with_activations:
+        raise ValueError("No pairs provided")
+    
+    # Extract activations by layer
+    pos_by_layer: Dict[int, List[torch.Tensor]] = {}
+    neg_by_layer: Dict[int, List[torch.Tensor]] = {}
+    
+    for pair in pairs_with_activations:
+        pos_acts = pair.positive_response.layers_activations
+        neg_acts = pair.negative_response.layers_activations
+        
+        for layer_key, act in pos_acts.items():
+            layer = int(layer_key)
+            if layers is None or layer in layers:
+                if layer not in pos_by_layer:
+                    pos_by_layer[layer] = []
+                pos_by_layer[layer].append(act.float() if act is not None else None)
+        
+        for layer_key, act in neg_acts.items():
+            layer = int(layer_key)
+            if layers is None or layer in layers:
+                if layer not in neg_by_layer:
+                    neg_by_layer[layer] = []
+                neg_by_layer[layer].append(act.float() if act is not None else None)
+    
+    # Stack into tensors
+    pos_tensors = {}
+    neg_tensors = {}
+    for layer in pos_by_layer:
+        valid_pos = [a for a in pos_by_layer[layer] if a is not None]
+        valid_neg = [a for a in neg_by_layer.get(layer, []) if a is not None]
+        if valid_pos and valid_neg:
+            pos_tensors[layer] = torch.stack(valid_pos)
+            neg_tensors[layer] = torch.stack(valid_neg)
+    
+    return detect_geometry_multi_layer(pos_tensors, neg_tensors, config)
+
+
+@dataclass
+class ExhaustiveCombinationResult:
+    """Result for a single layer combination."""
+    layers: Tuple[int, ...]
+    best_structure: StructureType
+    best_score: float
+    all_scores: Dict[str, float]
+
+
+@dataclass
+class ExhaustiveGeometryAnalysisResult:
+    """Results from exhaustive layer combination analysis."""
+    
+    total_combinations: int
+    """Total number of combinations tested."""
+    
+    all_results: List[ExhaustiveCombinationResult]
+    """All results, sorted by best_score descending."""
+    
+    best_combination: Tuple[int, ...]
+    """Layer combination with highest score."""
+    
+    best_score: float
+    """Highest score achieved."""
+    
+    best_structure: StructureType
+    """Structure type at best combination."""
+    
+    top_10: List[ExhaustiveCombinationResult]
+    """Top 10 combinations."""
+    
+    single_layer_best: int
+    """Best single layer."""
+    
+    single_layer_best_score: float
+    """Score of best single layer."""
+    
+    combination_beats_single: bool
+    """Whether any multi-layer combination beats best single layer."""
+    
+    improvement_over_single: float
+    """How much best combination improves over best single layer."""
+    
+    patterns: Dict[str, Any]
+    """Discovered patterns (layer frequency in top combinations, etc.)."""
+    
+    recommendation: str
+    """Final recommendation."""
+
+
+def detect_geometry_exhaustive(
+    pos_activations_by_layer: Dict[int, torch.Tensor],
+    neg_activations_by_layer: Dict[int, torch.Tensor],
+    max_layers: int = 16,
+    combination_method: str = "concat",
+    num_components: int = 5,
+    progress_callback: Optional[Callable[[int, int], None]] = None,
+    top_k: int = 100,
+) -> ExhaustiveGeometryAnalysisResult:
+    """
+    Exhaustively test all 2^N - 1 layer combinations for geometric structure.
+    
+    Memory-efficient: uses generators and only keeps top_k results in memory.
+    
+    Arguments:
+        pos_activations_by_layer: Dict mapping layer index to positive activations [N, hidden_dim]
+        neg_activations_by_layer: Dict mapping layer index to negative activations [N, hidden_dim]
+        max_layers: Maximum number of layers to consider (limits combinations)
+        combination_method: How to combine layers ("concat", "mean", "weighted")
+        num_components: Number of PCA components for analysis
+        progress_callback: Optional callback(current, total) for progress reporting
+        top_k: Number of top results to keep in memory (default 100)
+        
+    Returns:
+        ExhaustiveGeometryAnalysisResult with top combinations ranked
+    """
+    import heapq
+    from itertools import combinations as itertools_combinations
+    
+    layers = sorted(pos_activations_by_layer.keys())[:max_layers]
+    n_layers = len(layers)
+    
+    if n_layers == 0:
+        raise ValueError("No layers provided")
+    
+    geo_cfg = GeometryAnalysisConfig(num_components=num_components, optimization_steps=50)
+    
+    # Calculate total without building list (2^n - 1)
+    total_combinations = (1 << n_layers) - 1
+    
+    # Use min-heap to keep top_k results (negate scores for max-heap behavior)
+    top_results_heap: List[Tuple[float, ExhaustiveCombinationResult]] = []
+    single_layer_results: List[ExhaustiveCombinationResult] = []
+    
+    # Generator for combinations - no upfront memory allocation
+    def combo_generator():
+        for r in range(1, n_layers + 1):
+            for combo in itertools_combinations(layers, r):
+                yield combo
+    
+    # Test each combination
+    idx = 0
+    for combo in combo_generator():
+        idx += 1
+        if progress_callback:
+            progress_callback(idx, total_combinations)
+        
+        # Combine activations for this subset
+        if len(combo) == 1:
+            layer = combo[0]
+            combined_pos = pos_activations_by_layer[layer]
+            combined_neg = neg_activations_by_layer[layer]
+        else:
+            combined_pos, combined_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, 
+                list(combo), combination_method
+            )
+        
+        # Run geometry detection
+        result = detect_geometry_structure(combined_pos, combined_neg, geo_cfg)
+        
+        all_scores = {name: score.score for name, score in result.all_scores.items()}
+        combo_result = ExhaustiveCombinationResult(
+            layers=combo,
+            best_structure=result.best_structure,
+            best_score=result.best_score,
+            all_scores=all_scores,
+        )
+        
+        # Track single layer results separately
+        if len(combo) == 1:
+            single_layer_results.append(combo_result)
+        
+        # Maintain top_k using heap
+        if len(top_results_heap) < top_k:
+            heapq.heappush(top_results_heap, (combo_result.best_score, combo_result))
+        elif combo_result.best_score > top_results_heap[0][0]:
+            heapq.heapreplace(top_results_heap, (combo_result.best_score, combo_result))
+    
+    # Extract top results sorted by score descending
+    all_results = [r for _, r in sorted(top_results_heap, key=lambda x: -x[0])]
+    
+    # Extract insights
+    best_result = all_results[0] if all_results else None
+    best_combination = best_result.layers if best_result else ()
+    best_score = best_result.best_score if best_result else 0.0
+    best_structure = best_result.best_structure if best_result else StructureType.UNKNOWN
+    
+    top_10 = all_results[:10]
+    
+    # Find best single layer
+    if single_layer_results:
+        single_layer_results.sort(key=lambda x: x.best_score, reverse=True)
+        single_layer_best = single_layer_results[0].layers[0]
+        single_layer_best_score = single_layer_results[0].best_score
+    else:
+        single_layer_best = layers[0]
+        single_layer_best_score = 0.0
+    
+    combination_beats_single = best_score > single_layer_best_score
+    improvement_over_single = best_score - single_layer_best_score
+    
+    # Analyze patterns from top results
+    patterns = _analyze_combination_patterns(all_results, layers, top_k=min(50, len(all_results)))
+    
+    # Generate recommendation
+    recommendation = _generate_exhaustive_recommendation(
+        best_combination, best_score, best_structure,
+        single_layer_best, single_layer_best_score,
+        combination_beats_single, improvement_over_single,
+        patterns, total_combinations
+    )
+    
+    return ExhaustiveGeometryAnalysisResult(
+        total_combinations=total_combinations,
+        all_results=all_results,
+        best_combination=best_combination,
+        best_score=best_score,
+        best_structure=best_structure,
+        top_10=top_10,
+        single_layer_best=single_layer_best,
+        single_layer_best_score=single_layer_best_score,
+        combination_beats_single=combination_beats_single,
+        improvement_over_single=improvement_over_single,
+        patterns=patterns,
+        recommendation=recommendation,
+    )
+
+
+def detect_geometry_limited(
+    pos_activations_by_layer: Dict[int, torch.Tensor],
+    neg_activations_by_layer: Dict[int, torch.Tensor],
+    max_combo_size: int = 3,
+    combination_method: str = "concat",
+    num_components: int = 5,
+    progress_callback: Optional[Callable[[int, int], None]] = None,
+    top_k: int = 100,
+) -> ExhaustiveGeometryAnalysisResult:
+    """
+    Test limited layer combinations: 1-layer, 2-layer, ..., max_combo_size-layer, plus all layers.
+    
+    Much faster than exhaustive search while still finding good combinations.
+    For N layers with max_combo_size=3:
+    - 1-layer: N combinations
+    - 2-layer: N*(N-1)/2 combinations  
+    - 3-layer: N*(N-1)*(N-2)/6 combinations
+    - all-layers: 1 combination
+    
+    Total: O(N^3) instead of O(2^N)
+    
+    Arguments:
+        pos_activations_by_layer: Dict mapping layer index to positive activations [N, hidden_dim]
+        neg_activations_by_layer: Dict mapping layer index to negative activations [N, hidden_dim]
+        max_combo_size: Maximum combination size to test (1, 2, 3, etc.) before jumping to all
+        combination_method: How to combine layers ("concat", "mean", "weighted")
+        num_components: Number of PCA components for analysis
+        progress_callback: Optional callback(current, total) for progress reporting
+        top_k: Number of top results to keep in memory (default 100)
+        
+    Returns:
+        ExhaustiveGeometryAnalysisResult with top combinations ranked
+    """
+    import heapq
+    from itertools import combinations as itertools_combinations
+    from math import comb
+    
+    layers = sorted(pos_activations_by_layer.keys())
+    n_layers = len(layers)
+    
+    if n_layers == 0:
+        raise ValueError("No layers provided")
+    
+    geo_cfg = GeometryAnalysisConfig(num_components=num_components, optimization_steps=50)
+    
+    # Calculate total combinations: sum of C(n,r) for r=1 to max_combo_size, plus 1 for all layers
+    total_combinations = sum(comb(n_layers, r) for r in range(1, min(max_combo_size, n_layers) + 1))
+    if max_combo_size < n_layers:
+        total_combinations += 1  # Add all-layers combination
+    
+    # Use min-heap to keep top_k results
+    top_results_heap: List[Tuple[float, ExhaustiveCombinationResult]] = []
+    single_layer_results: List[ExhaustiveCombinationResult] = []
+    
+    # Generator for limited combinations
+    def combo_generator():
+        # 1-layer, 2-layer, ..., max_combo_size-layer
+        for r in range(1, min(max_combo_size, n_layers) + 1):
+            for combo in itertools_combinations(layers, r):
+                yield combo
+        # All layers (if not already included)
+        if max_combo_size < n_layers:
+            yield tuple(layers)
+    
+    # Test each combination
+    idx = 0
+    for combo in combo_generator():
+        idx += 1
+        if progress_callback:
+            progress_callback(idx, total_combinations)
+        
+        # Combine activations for this subset
+        if len(combo) == 1:
+            layer = combo[0]
+            combined_pos = pos_activations_by_layer[layer]
+            combined_neg = neg_activations_by_layer[layer]
+        else:
+            combined_pos, combined_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, 
+                list(combo), combination_method
+            )
+        
+        # Run geometry detection
+        result = detect_geometry_structure(combined_pos, combined_neg, geo_cfg)
+        
+        all_scores = {name: score.score for name, score in result.all_scores.items()}
+        combo_result = ExhaustiveCombinationResult(
+            layers=combo,
+            best_structure=result.best_structure,
+            best_score=result.best_score,
+            all_scores=all_scores,
+        )
+        
+        # Track single layer results separately
+        if len(combo) == 1:
+            single_layer_results.append(combo_result)
+        
+        # Maintain top_k using heap
+        if len(top_results_heap) < top_k:
+            heapq.heappush(top_results_heap, (combo_result.best_score, combo_result))
+        elif combo_result.best_score > top_results_heap[0][0]:
+            heapq.heapreplace(top_results_heap, (combo_result.best_score, combo_result))
+    
+    # Extract top results sorted by score descending
+    all_results = [r for _, r in sorted(top_results_heap, key=lambda x: -x[0])]
+    
+    # Extract insights
+    best_result = all_results[0] if all_results else None
+    best_combination = best_result.layers if best_result else ()
+    best_score = best_result.best_score if best_result else 0.0
+    best_structure = best_result.best_structure if best_result else StructureType.UNKNOWN
+    
+    top_10 = all_results[:10]
+    
+    # Find best single layer
+    if single_layer_results:
+        single_layer_results.sort(key=lambda x: x.best_score, reverse=True)
+        single_layer_best = single_layer_results[0].layers[0]
+        single_layer_best_score = single_layer_results[0].best_score
+    else:
+        single_layer_best = layers[0]
+        single_layer_best_score = 0.0
+    
+    combination_beats_single = best_score > single_layer_best_score
+    improvement_over_single = best_score - single_layer_best_score
+    
+    # Analyze patterns from top results
+    patterns = _analyze_combination_patterns(all_results, layers, top_k=min(50, len(all_results)))
+    
+    # Generate recommendation
+    recommendation = _generate_exhaustive_recommendation(
+        best_combination, best_score, best_structure,
+        single_layer_best, single_layer_best_score,
+        combination_beats_single, improvement_over_single,
+        patterns, total_combinations
+    )
+    
+    return ExhaustiveGeometryAnalysisResult(
+        total_combinations=total_combinations,
+        all_results=all_results,
+        best_combination=best_combination,
+        best_score=best_score,
+        best_structure=best_structure,
+        top_10=top_10,
+        single_layer_best=single_layer_best,
+        single_layer_best_score=single_layer_best_score,
+        combination_beats_single=combination_beats_single,
+        improvement_over_single=improvement_over_single,
+        patterns=patterns,
+        recommendation=recommendation,
+    )
+
+
+def detect_geometry_contiguous(
+    pos_activations_by_layer: Dict[int, torch.Tensor],
+    neg_activations_by_layer: Dict[int, torch.Tensor],
+    combination_method: str = "concat",
+    num_components: int = 5,
+    progress_callback: Optional[Callable[[int, int], None]] = None,
+    top_k: int = 100,
+) -> ExhaustiveGeometryAnalysisResult:
+    """
+    Test contiguous layer combinations only.
+    
+    Only tests combinations where layers are adjacent: 1-2, 2-3, 1-3, 5-8, etc.
+    Much faster: O(N^2) combinations instead of O(2^N).
+    
+    For N layers: N*(N+1)/2 combinations
+    - 36 layers: 666 combinations
+    - 24 layers: 300 combinations
+    
+    Arguments:
+        pos_activations_by_layer: Dict mapping layer index to positive activations [N, hidden_dim]
+        neg_activations_by_layer: Dict mapping layer index to negative activations [N, hidden_dim]
+        combination_method: How to combine layers ("concat", "mean", "weighted")
+        num_components: Number of PCA components for analysis
+        progress_callback: Optional callback(current, total) for progress reporting
+        top_k: Number of top results to keep in memory (default 100)
+        
+    Returns:
+        ExhaustiveGeometryAnalysisResult with top combinations ranked
+    """
+    import heapq
+    
+    layers = sorted(pos_activations_by_layer.keys())
+    n_layers = len(layers)
+    
+    if n_layers == 0:
+        raise ValueError("No layers provided")
+    
+    geo_cfg = GeometryAnalysisConfig(num_components=num_components, optimization_steps=50)
+    
+    # Total contiguous combinations: N*(N+1)/2
+    total_combinations = n_layers * (n_layers + 1) // 2
+    
+    # Use min-heap to keep top_k results
+    top_results_heap: List[Tuple[float, ExhaustiveCombinationResult]] = []
+    single_layer_results: List[ExhaustiveCombinationResult] = []
+    
+    # Generator for contiguous combinations
+    def combo_generator():
+        # For each starting layer
+        for start_idx in range(n_layers):
+            # For each ending layer (inclusive)
+            for end_idx in range(start_idx, n_layers):
+                yield tuple(layers[start_idx:end_idx + 1])
+    
+    # Test each combination
+    idx = 0
+    for combo in combo_generator():
+        idx += 1
+        if progress_callback:
+            progress_callback(idx, total_combinations)
+        
+        # Combine activations for this subset
+        if len(combo) == 1:
+            layer = combo[0]
+            combined_pos = pos_activations_by_layer[layer]
+            combined_neg = neg_activations_by_layer[layer]
+        else:
+            combined_pos, combined_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, 
+                list(combo), combination_method
+            )
+        
+        # Run geometry detection
+        result = detect_geometry_structure(combined_pos, combined_neg, geo_cfg)
+        
+        all_scores = {name: score.score for name, score in result.all_scores.items()}
+        combo_result = ExhaustiveCombinationResult(
+            layers=combo,
+            best_structure=result.best_structure,
+            best_score=result.best_score,
+            all_scores=all_scores,
+        )
+        
+        # Track single layer results separately
+        if len(combo) == 1:
+            single_layer_results.append(combo_result)
+        
+        # Maintain top_k using heap
+        if len(top_results_heap) < top_k:
+            heapq.heappush(top_results_heap, (combo_result.best_score, combo_result))
+        elif combo_result.best_score > top_results_heap[0][0]:
+            heapq.heapreplace(top_results_heap, (combo_result.best_score, combo_result))
+    
+    # Extract top results sorted by score descending
+    all_results = [r for _, r in sorted(top_results_heap, key=lambda x: -x[0])]
+    
+    # Extract insights
+    best_result = all_results[0] if all_results else None
+    best_combination = best_result.layers if best_result else ()
+    best_score = best_result.best_score if best_result else 0.0
+    best_structure = best_result.best_structure if best_result else StructureType.UNKNOWN
+    
+    top_10 = all_results[:10]
+    
+    # Find best single layer
+    if single_layer_results:
+        single_layer_results.sort(key=lambda x: x.best_score, reverse=True)
+        single_layer_best = single_layer_results[0].layers[0]
+        single_layer_best_score = single_layer_results[0].best_score
+    else:
+        single_layer_best = layers[0]
+        single_layer_best_score = 0.0
+    
+    combination_beats_single = best_score > single_layer_best_score
+    improvement_over_single = best_score - single_layer_best_score
+    
+    # Analyze patterns from top results
+    patterns = _analyze_combination_patterns(all_results, layers, top_k=min(50, len(all_results)))
+    
+    # Generate recommendation
+    recommendation = _generate_exhaustive_recommendation(
+        best_combination, best_score, best_structure,
+        single_layer_best, single_layer_best_score,
+        combination_beats_single, improvement_over_single,
+        patterns, total_combinations
+    )
+    
+    return ExhaustiveGeometryAnalysisResult(
+        total_combinations=total_combinations,
+        all_results=all_results,
+        best_combination=best_combination,
+        best_score=best_score,
+        best_structure=best_structure,
+        top_10=top_10,
+        single_layer_best=single_layer_best,
+        single_layer_best_score=single_layer_best_score,
+        combination_beats_single=combination_beats_single,
+        improvement_over_single=improvement_over_single,
+        patterns=patterns,
+        recommendation=recommendation,
+    )
+
+
+def detect_geometry_smart(
+    pos_activations_by_layer: Dict[int, torch.Tensor],
+    neg_activations_by_layer: Dict[int, torch.Tensor],
+    max_combo_size: int = 3,
+    combination_method: str = "concat",
+    num_components: int = 5,
+    progress_callback: Optional[Callable[[int, int], None]] = None,
+    top_k: int = 100,
+) -> ExhaustiveGeometryAnalysisResult:
+    """
+    Smart layer combination search: contiguous + limited (1,2,3-layer) combinations.
+    
+    Tests:
+    1. All contiguous combinations (L1-L2, L1-L3, L5-L10, etc.)
+    2. All 1,2,3-layer non-contiguous combinations
+    
+    Deduplicates overlapping combinations.
+    
+    For N=36 layers with max_combo_size=3:
+    - Contiguous: 666 combinations
+    - Limited non-contiguous: ~7,100 additional combinations
+    - Total: ~7,800 unique combinations
+    
+    Arguments:
+        pos_activations_by_layer: Dict mapping layer index to positive activations [N, hidden_dim]
+        neg_activations_by_layer: Dict mapping layer index to negative activations [N, hidden_dim]
+        max_combo_size: Maximum combination size for non-contiguous (default: 3)
+        combination_method: How to combine layers ("concat", "mean", "weighted")
+        num_components: Number of PCA components for analysis
+        progress_callback: Optional callback(current, total) for progress reporting
+        top_k: Number of top results to keep in memory (default 100)
+        
+    Returns:
+        ExhaustiveGeometryAnalysisResult with top combinations ranked
+    """
+    import heapq
+    from itertools import combinations as itertools_combinations
+    
+    layers = sorted(pos_activations_by_layer.keys())
+    n_layers = len(layers)
+    
+    if n_layers == 0:
+        raise ValueError("No layers provided")
+    
+    geo_cfg = GeometryAnalysisConfig(num_components=num_components, optimization_steps=50)
+    
+    # Generate all unique combinations: contiguous + limited
+    all_combos_set: set = set()
+    
+    # Add contiguous combinations
+    for start_idx in range(n_layers):
+        for end_idx in range(start_idx, n_layers):
+            all_combos_set.add(tuple(layers[start_idx:end_idx + 1]))
+    
+    # Add limited combinations (1,2,3-layer)
+    for r in range(1, min(max_combo_size, n_layers) + 1):
+        for combo in itertools_combinations(layers, r):
+            all_combos_set.add(combo)
+    
+    # Convert to sorted list
+    all_combos = sorted(all_combos_set, key=lambda x: (len(x), x))
+    total_combinations = len(all_combos)
+    
+    # Use min-heap to keep top_k results
+    top_results_heap: List[Tuple[float, ExhaustiveCombinationResult]] = []
+    single_layer_results: List[ExhaustiveCombinationResult] = []
+    
+    # Test each combination
+    for idx, combo in enumerate(all_combos):
+        if progress_callback:
+            progress_callback(idx + 1, total_combinations)
+        
+        # Combine activations for this subset
+        if len(combo) == 1:
+            layer = combo[0]
+            combined_pos = pos_activations_by_layer[layer]
+            combined_neg = neg_activations_by_layer[layer]
+        else:
+            combined_pos, combined_neg = _combine_layer_activations(
+                pos_activations_by_layer, neg_activations_by_layer, 
+                list(combo), combination_method
+            )
+        
+        # Run geometry detection
+        result = detect_geometry_structure(combined_pos, combined_neg, geo_cfg)
+        
+        all_scores = {name: score.score for name, score in result.all_scores.items()}
+        combo_result = ExhaustiveCombinationResult(
+            layers=combo,
+            best_structure=result.best_structure,
+            best_score=result.best_score,
+            all_scores=all_scores,
+        )
+        
+        # Track single layer results separately
+        if len(combo) == 1:
+            single_layer_results.append(combo_result)
+        
+        # Maintain top_k using heap
+        if len(top_results_heap) < top_k:
+            heapq.heappush(top_results_heap, (combo_result.best_score, combo_result))
+        elif combo_result.best_score > top_results_heap[0][0]:
+            heapq.heapreplace(top_results_heap, (combo_result.best_score, combo_result))
+    
+    # Extract top results sorted by score descending
+    all_results = [r for _, r in sorted(top_results_heap, key=lambda x: -x[0])]
+    
+    # Extract insights
+    best_result = all_results[0] if all_results else None
+    best_combination = best_result.layers if best_result else ()
+    best_score = best_result.best_score if best_result else 0.0
+    best_structure = best_result.best_structure if best_result else StructureType.UNKNOWN
+    
+    top_10 = all_results[:10]
+    
+    # Find best single layer
+    if single_layer_results:
+        single_layer_results.sort(key=lambda x: x.best_score, reverse=True)
+        single_layer_best = single_layer_results[0].layers[0]
+        single_layer_best_score = single_layer_results[0].best_score
+    else:
+        single_layer_best = layers[0]
+        single_layer_best_score = 0.0
+    
+    combination_beats_single = best_score > single_layer_best_score
+    improvement_over_single = best_score - single_layer_best_score
+    
+    # Analyze patterns from top results
+    patterns = _analyze_combination_patterns(all_results, layers, top_k=min(50, len(all_results)))
+    
+    # Generate recommendation
+    recommendation = _generate_exhaustive_recommendation(
+        best_combination, best_score, best_structure,
+        single_layer_best, single_layer_best_score,
+        combination_beats_single, improvement_over_single,
+        patterns, total_combinations
+    )
+    
+    return ExhaustiveGeometryAnalysisResult(
+        total_combinations=total_combinations,
+        all_results=all_results,
+        best_combination=best_combination,
+        best_score=best_score,
+        best_structure=best_structure,
+        top_10=top_10,
+        single_layer_best=single_layer_best,
+        single_layer_best_score=single_layer_best_score,
+        combination_beats_single=combination_beats_single,
+        improvement_over_single=improvement_over_single,
+        patterns=patterns,
+        recommendation=recommendation,
+    )
+
+
+def _analyze_combination_patterns(
+    all_results: List[ExhaustiveCombinationResult],
+    layers: List[int],
+    top_k: int = 50,
+) -> Dict[str, Any]:
+    """Analyze patterns in top combinations."""
+    from collections import Counter
+    
+    top_results = all_results[:top_k]
+    
+    # Layer frequency in top combinations
+    layer_freq = Counter()
+    for r in top_results:
+        for layer in r.layers:
+            layer_freq[layer] += 1
+    
+    # Combination size distribution in top results
+    size_dist = Counter(len(r.layers) for r in top_results)
+    
+    # Best score by combination size
+    size_to_best: Dict[int, float] = {}
+    for r in all_results:
+        size = len(r.layers)
+        if size not in size_to_best or r.best_score > size_to_best[size]:
+            size_to_best[size] = r.best_score
+    
+    # Structure frequency in top combinations
+    structure_freq = Counter(r.best_structure for r in top_results)
+    
+    # Adjacent layer pairs in top combinations
+    adjacent_count = 0
+    for r in top_results:
+        if len(r.layers) >= 2:
+            sorted_layers = sorted(r.layers)
+            for i in range(len(sorted_layers) - 1):
+                if sorted_layers[i + 1] - sorted_layers[i] == 1:
+                    adjacent_count += 1
+                    break
+    
+    # Layer position analysis (early vs late layers)
+    mid_layer = layers[len(layers) // 2] if layers else 0
+    early_in_top = sum(1 for r in top_results for l in r.layers if l < mid_layer)
+    late_in_top = sum(1 for r in top_results for l in r.layers if l >= mid_layer)
+    
+    return {
+        "layer_frequency_in_top": dict(layer_freq.most_common()),
+        "most_important_layers": [l for l, _ in layer_freq.most_common(5)],
+        "size_distribution_in_top": dict(size_dist),
+        "best_score_by_size": size_to_best,
+        "optimal_combination_size": max(size_to_best.keys(), key=lambda k: size_to_best[k]) if size_to_best else 1,
+        "structure_frequency_in_top": {s.value: c for s, c in structure_freq.most_common()},
+        "dominant_structure": structure_freq.most_common(1)[0][0].value if structure_freq else "unknown",
+        "adjacent_pairs_in_top": adjacent_count,
+        "early_vs_late_ratio": early_in_top / late_in_top if late_in_top > 0 else float('inf'),
+    }
+
+
+def _generate_exhaustive_recommendation(
+    best_combination: Tuple[int, ...],
+    best_score: float,
+    best_structure: StructureType,
+    single_layer_best: int,
+    single_layer_best_score: float,
+    combination_beats_single: bool,
+    improvement_over_single: float,
+    patterns: Dict[str, Any],
+    total_combinations: int,
+) -> str:
+    """Generate recommendation from exhaustive analysis."""
+    parts = []
+    
+    parts.append(f"Tested {total_combinations} layer combinations.")
+    
+    if combination_beats_single and improvement_over_single > 0.05:
+        layers_str = "+".join(f"L{l}" for l in best_combination)
+        parts.append(
+            f"BEST: {layers_str} ({best_structure.value}: {best_score:.3f}), "
+            f"+{improvement_over_single:.3f} over single layer L{single_layer_best}."
+        )
+    else:
+        parts.append(
+            f"BEST: Single layer L{single_layer_best} ({best_score:.3f}). "
+            f"Multi-layer combinations don't significantly improve."
+        )
+    
+    # Pattern insights
+    opt_size = patterns.get("optimal_combination_size", 1)
+    if opt_size > 1:
+        parts.append(f"Optimal combination size: {opt_size} layers.")
+    
+    important_layers = patterns.get("most_important_layers", [])
+    if important_layers:
+        layers_str = ", ".join(f"L{l}" for l in important_layers[:3])
+        parts.append(f"Most important layers: {layers_str}.")
+    
+    dominant = patterns.get("dominant_structure", "unknown")
+    parts.append(f"Dominant structure: {dominant}.")
+    
+    return " ".join(parts)