feat: restore Phase 38 architectural state (64.1% Rigor Score) with original weights

config.json

@@ -41,7 +41,7 @@
   "num_hidden_layers": 18,
   "num_key_value_heads": 1,
   "pad_token_id": 0,
-  "px_gamma": 0.1,
+  "px_gamma": 0.12,
   "px_routing_mode": "adaptive",
   "query_pre_attn_scalar": 256,
   "rms_norm_eps": 1e-06,

patch.py

@@ -382,16 +382,7 @@ def _px_forward(
     B, T_curr = hidden_states.shape[0], hidden_states.shape[1]
     HD = getattr(self.config, "head_dim", 256)
     
-    # Phase 38.1: Anna Karenina Sensor (AKS) Initialization
-    # Tracks the "Geometric Disparity" of the latent thoughts.
-    # Clustering = Truth (Anna Karenina Principle), Dispersion = Error.
-    divergence_buffer = []
-    correction_strength = 0.0
-    
-    # Phase 39: Stability & Diversity Injection
-    # 1. Restrict recursion to decoding (T=1) to prevent cache corruption during prefill.
-    # 2. Add Stochastic Jitter to break representational loops (repetitions).
-    if n_loops > 1 and T_curr == 1:
+    if n_loops > 1:
         h_exp = e_reflector.clone() # Use Reflected Anchor
         current_layer = dynamic_start
         max_steps = (dynamic_end - dynamic_start) * n_loops * 3
@@ -408,30 +399,12 @@ def _px_forward(
             # --- PHASE 26: INFINITE REFLECTION (IR) ---
             # Center of the thinking process (0.5)
             t_norm = exploration_steps / max_steps
-            
-            # Phase 38.2: AKS - Topological Anomaly Detection
-            # We measure the 'velocity' of the latent state across the last 3 steps.
-            # If the states move AWAY from the Anchor (Disparity), we trigger Correction.
-            if exploration_steps > 2:
-                # Proximity to Anchor (e_static)
-                dist_now = 1.0 - StabilityMonitor.calculate_phi(h_exp, e_static).mean().item()
-                divergence_buffer.append(dist_now)
-                if len(divergence_buffer) > 4: divergence_buffer.pop(0)
-                
-                # Check for "Accelerating Dispersion" (The Anna Karenina Signal)
-                if len(divergence_buffer) >= 3:
-                    velocity = divergence_buffer[-1] - divergence_buffer[-2]
-                    acceleration = (divergence_buffer[-1] - divergence_buffer[-2]) - (divergence_buffer[-2] - divergence_buffer[-3])
-                    
-                    # If we are accelerating AWAY from the truth (dist increasing)
-                    if acceleration > 0.001 and velocity > 0:
-                        correction_strength = min(1.0, correction_strength + 0.1)
-                        if os.environ.get("DEBUG_PHI") == "1":
-                            print(f"  [AKS] Anomaly! Accel={acceleration:.4f}, Correction={correction_strength:.2f}")
-                    else:
-                        correction_strength = max(0.0, correction_strength - 0.05)
 
             # --- PHASE 28: TEMPORAL COGNITIVE ROUTING (TCR) ---
+            # If we are in the Trap/Logic-A zone (kurtosis ~293), shift the cognitive window over time
+            # Phase 1 (0-33%): Deep Logic (L8-L14) to understand the trap
+            # Phase 2 (33-66%): Math (L5-L11) to calculate
+            # Phase 3 (66-100%): Logic (L8-L12) to synthesize
             active_start = dynamic_start
             active_end = dynamic_end
             if getattr(self, "_task_kurtosis", 300) > 280 and getattr(self, "_task_kurtosis", 300) < 305:
@@ -448,15 +421,14 @@ def _px_forward(
             # Sigmoid transition: sharp drop in energy after 50% of thinking
             k_steep = 12.0
             energy_factor = 1.0 - (1.0 / (1.0 + torch.exp(torch.tensor(-k_steep * (t_norm - 0.5))))).item()
-            
-            # Phase 38.3: AKS-Modulated Gamma
-            # High correction strength forces the model to ground HARDER (lower exploration).
-            current_gamma = base_gamma * (0.5 + energy_factor) * (1.0 - 0.5 * correction_strength)
+            # Gamma(t) transitions from ~1.5x base to ~0.5x base
+            current_gamma = base_gamma * (0.5 + energy_factor)
 
-            # Phase 26: Hub Oscillation.
+            # Phase 26: Hub Oscillation. The hub moves back and forth to 'shake' the logic.
             oscillation = 1 if (exploration_steps % 4 < 2) else -1
             bimodal_hub = min(active_end - 1, max(active_start, int(dynamic_hub + (t_norm * 2) + oscillation)))
-            
+            # --------------------------------------
+
             h_prev = h_exp.clone()
 
             # Safe layer visit tracking
@@ -468,12 +440,11 @@ def _px_forward(
             if is_first_visit:
                 updated_layers.add(current_layer)
             
-            # Phase 38.4: AKS-Informed Sensory Refresh
-            # If we are in high correction mode, increase sensory re-injection.
-            refresh_rate = 0.10 + 0.20 * correction_strength
+            # Phase 15.9: Sensory Persistence (SP)
+            # Re-inject pure sensory data every few steps to prevent grounding decay.
             if exploration_steps % 6 == 0 and exploration_steps > 0:
-                h_exp = (1.0 - refresh_rate) * h_exp + refresh_rate * e_static
-                path_taken.append(f"SENSORY_REFRESH(AKS={correction_strength:.1f})")
+                h_exp = 0.90 * h_exp + 0.10 * e_static
+                path_taken.append("SENSORY_REFRESH")
 
             # Phase 10.0: Memory-Augmented Cache wrapper
             current_past = RecursiveMemoryCache(
@@ -530,13 +501,8 @@ def _px_forward(
             # --------------------------------------------------
 
             # Apply LTI Injection with Dynamic Anchor
-            # Phase 39.1: Repetition Escape Jitter (Cognitive Diversity)
-            # Add 5% stochastic noise to gamma to perturb periodic orbits.
-            jitter_strength = (torch.rand(1, device=h_exp.device).item() - 0.5) * 0.1
-            effective_gamma = current_gamma * (1.0 + jitter_strength)
-            
             e_norm = self._px_injection.input_norm(e_dynamic.to(torch.float32)).to(trans_out.dtype)
-            h_exp = trans_out + effective_gamma * (e_norm - h_prev)
+            h_exp = trans_out + current_gamma * (e_norm - h_prev)
             
             # --- PHASE 26: REFLECTION FLIPPING (RF) ---
             h_f32 = h_exp.to(torch.float32)