diff --git "a/axquant_plan.json" "b/axquant_plan.json"
--- "a/axquant_plan.json"
+++ "b/axquant_plan.json"
@@ -1,8 +1,8 @@
 {
-  "analysis_sha256": "adcabd6cfe28d52f7d575002c4dd9598af1bf7ff2302d2573a6461aa4a70f623",
+  "analysis_sha256": "970a8f05f20eb0bb2df3ab73cdd8dd5fb31df2de312257c1beb2e335f9c8660b",
   "architecture_profile": {
     "adapter_id": "qwen36-v1",
-    "config_model_type": "qwen3_5",
+    "config_model_type": "qwen3_5_moe",
     "dense": false,
     "mtp_declared": true,
     "notes": [
@@ -1005,31 +1005,31 @@
       "tensor": "model.language_model.layers.0.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 64,
+      "bits": 8,
+      "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.04275,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.07600000000000001,
-        "long_context_loss": 0.028499999999999998,
-        "mtp_acceptance_loss": 0.0076,
-        "output_kl": 0.095,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.061750000000000006,
-        "token_disagreement": 0.052250000000000005
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.0.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.05187,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 6.5
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.0.mlp.shared_expert_gate.weight"
     },
@@ -1382,31 +1382,31 @@
       "tensor": "model.language_model.layers.1.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 64,
+      "bits": 8,
+      "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.04275,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.07600000000000001,
-        "long_context_loss": 0.028499999999999998,
-        "mtp_acceptance_loss": 0.0076,
-        "output_kl": 0.095,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.061750000000000006,
-        "token_disagreement": 0.052250000000000005
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.1.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.05187,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 6.5
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.1.mlp.shared_expert_gate.weight"
     },
@@ -11213,31 +11213,31 @@
       "tensor": "model.language_model.layers.2.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 64,
+      "bits": 8,
+      "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.04275,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.07600000000000001,
-        "long_context_loss": 0.028499999999999998,
-        "mtp_acceptance_loss": 0.0076,
-        "output_kl": 0.095,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.061750000000000006,
-        "token_disagreement": 0.052250000000000005
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.2.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.05187,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 6.5
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.2.mlp.shared_expert_gate.weight"
     },
@@ -11474,31 +11474,31 @@
       "tensor": "model.language_model.layers.3.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.3.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.3.mlp.shared_expert_gate.weight"
     },
@@ -12170,31 +12170,31 @@
       "tensor": "model.language_model.layers.4.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.4.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.4.mlp.shared_expert_gate.weight"
     },
@@ -12692,31 +12692,31 @@
       "tensor": "model.language_model.layers.5.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.5.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.5.mlp.shared_expert_gate.weight"
     },
@@ -13214,31 +13214,31 @@
       "tensor": "model.language_model.layers.6.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.6.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.6.mlp.shared_expert_gate.weight"
     },
@@ -13475,31 +13475,31 @@
       "tensor": "model.language_model.layers.7.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.7.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.7.mlp.shared_expert_gate.weight"
     },
@@ -14461,31 +14461,31 @@
       "tensor": "model.language_model.layers.10.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.10.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.10.mlp.shared_expert_gate.weight"
     },
@@ -14838,31 +14838,31 @@
       "tensor": "model.language_model.layers.8.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.8.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.8.mlp.shared_expert_gate.weight"
     },
@@ -15215,31 +15215,31 @@
       "tensor": "model.language_model.layers.9.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.9.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.9.mlp.shared_expert_gate.weight"
     },
@@ -15853,31 +15853,31 @@
       "tensor": "model.language_model.layers.11.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.11.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.11.mlp.shared_expert_gate.weight"
     },
@@ -16317,31 +16317,31 @@
       "tensor": "model.language_model.layers.12.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.12.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.12.mlp.shared_expert_gate.weight"
     },
@@ -16694,31 +16694,31 @@
       "tensor": "model.language_model.layers.13.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.13.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.13.mlp.shared_expert_gate.weight"
     },
@@ -17303,31 +17303,31 @@
       "tensor": "model.language_model.layers.14.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.14.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.14.linear_attn.in_proj_b.weight"
     },
@@ -17477,31 +17477,31 @@
       "tensor": "model.language_model.layers.14.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.14.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.14.mlp.shared_expert_gate.weight"
     },
@@ -17709,31 +17709,31 @@
       "tensor": "model.language_model.layers.15.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.15.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.15.mlp.shared_expert_gate.weight"
     },
@@ -17941,60 +17941,60 @@
       "tensor": "model.language_model.layers.16.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.16.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.16.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.16.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.16.linear_attn.in_proj_b.weight"
     },
@@ -18173,31 +18173,31 @@
       "tensor": "model.language_model.layers.16.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.16.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.16.mlp.shared_expert_gate.weight"
     },
@@ -18753,60 +18753,60 @@
       "tensor": "model.language_model.layers.17.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.17.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.17.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.17.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.17.linear_attn.in_proj_b.weight"
     },
@@ -18956,31 +18956,31 @@
       "tensor": "model.language_model.layers.17.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.17.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.17.mlp.shared_expert_gate.weight"
     },
@@ -19130,60 +19130,60 @@
       "tensor": "model.language_model.layers.18.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.18.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.18.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.18.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.18.linear_attn.in_proj_b.weight"
     },
@@ -19362,31 +19362,31 @@
       "tensor": "model.language_model.layers.18.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.18.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.18.mlp.shared_expert_gate.weight"
     },
@@ -19594,31 +19594,31 @@
       "tensor": "model.language_model.layers.19.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.19.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.19.mlp.shared_expert_gate.weight"
     },
@@ -20203,60 +20203,60 @@
       "tensor": "model.language_model.layers.20.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.20.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.20.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.20.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.20.linear_attn.in_proj_b.weight"
     },
@@ -20406,31 +20406,31 @@
       "tensor": "model.language_model.layers.20.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.20.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.20.mlp.shared_expert_gate.weight"
     },
@@ -20580,60 +20580,60 @@
       "tensor": "model.language_model.layers.21.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.21.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.21.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.21.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.21.linear_attn.in_proj_b.weight"
     },
@@ -20812,31 +20812,31 @@
       "tensor": "model.language_model.layers.21.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.21.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.21.mlp.shared_expert_gate.weight"
     },
@@ -20986,60 +20986,60 @@
       "tensor": "model.language_model.layers.22.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.22.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.22.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.22.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.22.linear_attn.in_proj_b.weight"
     },
@@ -21218,31 +21218,31 @@
       "tensor": "model.language_model.layers.22.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.22.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.22.mlp.shared_expert_gate.weight"
     },
@@ -21334,31 +21334,31 @@
       "tensor": "model.language_model.layers.23.mlp.experts.gate_up_proj"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.23.self_attn.k_proj",
       "outlier_strategy": "none",
       "parameters": 1048576,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.23.self_attn.k_proj.weight"
     },
@@ -21421,31 +21421,31 @@
       "tensor": "model.language_model.layers.23.self_attn.q_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.23.self_attn.v_proj",
       "outlier_strategy": "none",
       "parameters": 1048576,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.23.self_attn.v_proj.weight"
     },
@@ -21856,31 +21856,31 @@
       "tensor": "model.language_model.layers.23.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.23.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.23.mlp.shared_expert_gate.weight"
     },
@@ -22088,60 +22088,60 @@
       "tensor": "model.language_model.layers.24.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.24.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.24.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.24.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.24.linear_attn.in_proj_b.weight"
     },
@@ -22291,31 +22291,31 @@
       "tensor": "model.language_model.layers.24.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.24.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.24.mlp.shared_expert_gate.weight"
     },
@@ -22465,60 +22465,60 @@
       "tensor": "model.language_model.layers.25.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.25.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.25.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.25.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.25.linear_attn.in_proj_b.weight"
     },
@@ -22697,31 +22697,31 @@
       "tensor": "model.language_model.layers.25.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.25.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.25.mlp.shared_expert_gate.weight"
     },
@@ -22900,31 +22900,31 @@
       "tensor": "model.language_model.layers.27.mlp.experts.gate_up_proj"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.27.self_attn.k_proj",
       "outlier_strategy": "none",
       "parameters": 1048576,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.27.self_attn.k_proj.weight"
     },
@@ -22987,31 +22987,31 @@
       "tensor": "model.language_model.layers.27.self_attn.q_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.27.self_attn.v_proj",
       "outlier_strategy": "none",
       "parameters": 1048576,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.27.self_attn.v_proj.weight"
     },
@@ -23277,60 +23277,60 @@
       "tensor": "model.language_model.layers.26.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.26.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.26.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.26.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.26.linear_attn.in_proj_b.weight"
     },
@@ -23509,31 +23509,31 @@
       "tensor": "model.language_model.layers.26.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.26.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.26.mlp.shared_expert_gate.weight"
     },
@@ -23741,31 +23741,31 @@
       "tensor": "model.language_model.layers.27.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.27.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.27.mlp.shared_expert_gate.weight"
     },
@@ -23973,60 +23973,60 @@
       "tensor": "model.language_model.layers.28.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.28.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.28.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.28.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.28.linear_attn.in_proj_b.weight"
     },
@@ -24176,31 +24176,31 @@
       "tensor": "model.language_model.layers.28.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.28.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.28.mlp.shared_expert_gate.weight"
     },
@@ -24524,31 +24524,31 @@
       "tensor": "model.language_model.layers.31.mlp.experts.gate_up_proj"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.31.self_attn.k_proj",
       "outlier_strategy": "none",
       "parameters": 1048576,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.31.self_attn.k_proj.weight"
     },
@@ -24611,31 +24611,31 @@
       "tensor": "model.language_model.layers.31.self_attn.q_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.31.self_attn.v_proj",
       "outlier_strategy": "none",
       "parameters": 1048576,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.31.self_attn.v_proj.weight"
     },
@@ -24756,60 +24756,60 @@
       "tensor": "model.language_model.layers.29.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.29.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.29.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.29.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.29.linear_attn.in_proj_b.weight"
     },
@@ -24988,31 +24988,31 @@
       "tensor": "model.language_model.layers.29.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.29.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.29.mlp.shared_expert_gate.weight"
     },
@@ -25162,60 +25162,60 @@
       "tensor": "model.language_model.layers.30.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.30.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.30.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.30.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.30.linear_attn.in_proj_b.weight"
     },
@@ -25394,31 +25394,31 @@
       "tensor": "model.language_model.layers.30.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.30.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.30.mlp.shared_expert_gate.weight"
     },
@@ -25597,31 +25597,31 @@
       "tensor": "model.language_model.layers.31.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.31.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.31.mlp.shared_expert_gate.weight"
     },
@@ -25829,60 +25829,60 @@
       "tensor": "model.language_model.layers.32.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.32.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.32.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.32.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.32.linear_attn.in_proj_b.weight"
     },
@@ -26177,31 +26177,31 @@
       "tensor": "model.language_model.layers.32.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.32.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.32.mlp.shared_expert_gate.weight"
     },
@@ -26351,60 +26351,60 @@
       "tensor": "model.language_model.layers.33.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.33.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.33.linear_attn.in_proj_a.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.33.linear_attn.in_proj_b",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.33.linear_attn.in_proj_b.weight"
     },
@@ -26699,31 +26699,31 @@
       "tensor": "model.language_model.layers.33.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.33.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.33.mlp.shared_expert_gate.weight"
     },
@@ -26873,31 +26873,31 @@
       "tensor": "model.language_model.layers.34.linear_attn.dt_bias"
     },
     {
-      "bits": 6,
-      "group_size": 32,
+      "bits": 8,
+      "group_size": 64,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.057275649276110355,
+        "cosine_distance": 0.02025,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.10182337649086284,
-        "long_context_loss": 0.15909902576697318,
-        "mtp_acceptance_loss": 0.010182337649086284,
-        "output_kl": 0.12727922061357855,
+        "hidden_state_error": 0.036,
+        "long_context_loss": 0.056249999999999994,
+        "mtp_acceptance_loss": 0.0036,
+        "output_kl": 0.045,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.08273149339882606,
-        "token_disagreement": 0.0700035713374682
+        "task_loss_delta": 0.029249999999999998,
+        "token_disagreement": 0.02475
       },
       "module_path": "model.language_model.layers.34.linear_attn.in_proj_a",
       "outlier_strategy": "none",
       "parameters": 65536,
-      "predicted_loss": 0.07916767522164586,
+      "predicted_loss": 0.027989999999999998,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "storage_bpw": 8.5
       },
       "tensor": "model.language_model.layers.34.linear_attn.in_proj_a.weight"
     },
@@ -26960,31 +26960,31 @@
       "tensor": "model.language_model.layers.34.linear_attn.in_proj_qkv.weight"
     },
     {
-      "bits": 8,
-      "group_size": 64,
+      "bits": 6,
+      "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.02025,
+        "cosine_distance": 0.057275649276110355,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.036,
-        "long_context_loss": 0.056249999999999994,
-        "mtp_acceptance_loss": 0.0036,
-        "output_kl": 0.045,
+        "hidden_state_error": 0.10182337649086284,
+        "long_context_loss": 0.15909902576697318,
+        "mtp_acceptance_loss": 0.010182337649086284,
+        "output_kl": 0.12727922061357855,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.029249999999999998,
-        "token_disagreement": 0.02475
+        "task_loss_delta": 0.08273149339882606,
+        "token_disagreement": 0.0700035713374682
       },
       "module_path": "model.language_model.layers.34.linear_attn.in_proj_z",
       "outlier_strategy": "none",
       "parameters": 8388608,
-      "predicted_loss": 0.027989999999999998,
+      "predicted_loss": 0.07916767522164586,
       "reason": "selected by marginal quality gain per storage bit",
       "role": "attention",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
         "selected_from_candidates": 7,
-        "storage_bpw": 8.5
+        "storage_bpw": 7.0
       },
       "tensor": "model.language_model.layers.34.linear_attn.in_proj_z.weight"
     },
@@ -27221,31 +27221,31 @@
       "tensor": "model.language_model.layers.34.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.34.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.34.mlp.shared_expert_gate.weight"
     },
@@ -27482,31 +27482,31 @@
       "tensor": "model.language_model.layers.35.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.35.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.35.mlp.shared_expert_gate.weight"
     },
@@ -28178,31 +28178,31 @@
       "tensor": "model.language_model.layers.36.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.36.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.36.mlp.shared_expert_gate.weight"
     },
@@ -28700,31 +28700,31 @@
       "tensor": "model.language_model.layers.37.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
+      "bits": 8,
       "group_size": 32,
       "method": "affine",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.021876116042958818,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.03889087296526012,
+        "long_context_loss": 0.014584077361972545,
+        "mtp_acceptance_loss": 0.003889087296526012,
+        "output_kl": 0.04861359120657515,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.03159883428427385,
+        "token_disagreement": 0.026737475163616336
       },
       "module_path": "model.language_model.layers.37.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
+      "predicted_loss": 0.02654302079879003,
+      "reason": "protected router policy",
+      "role": "router",
       "scale_strategy": "group-affine",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 9.0
       },
       "tensor": "model.language_model.layers.37.mlp.shared_expert_gate.weight"
     },
@@ -29222,31 +29222,31 @@
       "tensor": "model.language_model.layers.38.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
-      "method": "affine",
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.0,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
       },
       "module_path": "model.language_model.layers.38.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
-      "scale_strategy": "group-affine",
+      "predicted_loss": 0.0,
+      "reason": "protected router policy",
+      "role": "router",
+      "scale_strategy": "none",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 16.0
       },
       "tensor": "model.language_model.layers.38.mlp.shared_expert_gate.weight"
     },
@@ -29541,31 +29541,31 @@
       "tensor": "model.language_model.layers.39.mlp.shared_expert.up_proj.weight"
     },
     {
-      "bits": 6,
-      "group_size": 32,
-      "method": "affine",
+      "bits": 16,
+      "group_size": null,
+      "method": "bf16",
       "metrics": {
-        "cosine_distance": 0.03022881489572491,
+        "cosine_distance": 0.0,
         "decode_latency_cost": 0.0,
-        "hidden_state_error": 0.05374011537017762,
-        "long_context_loss": 0.020152543263816602,
-        "mtp_acceptance_loss": 0.005374011537017762,
-        "output_kl": 0.06717514421272203,
+        "hidden_state_error": 0.0,
+        "long_context_loss": 0.0,
+        "mtp_acceptance_loss": 0.0,
+        "output_kl": 0.0,
         "peak_memory_cost": 0.0,
         "prefill_latency_cost": 0.0,
-        "task_loss_delta": 0.043663843738269315,
-        "token_disagreement": 0.036946329316997116
+        "task_loss_delta": 0.0,
+        "token_disagreement": 0.0
       },
       "module_path": "model.language_model.layers.39.mlp.shared_expert_gate",
       "outlier_strategy": "none",
       "parameters": 2048,
-      "predicted_loss": 0.036677628740146226,
-      "reason": "selected by marginal quality gain per storage bit",
-      "role": "expert",
-      "scale_strategy": "group-affine",
+      "predicted_loss": 0.0,
+      "reason": "protected router policy",
+      "role": "router",
+      "scale_strategy": "none",
       "strategy_metadata": {
-        "selected_from_candidates": 7,
-        "storage_bpw": 7.0
+        "selected_from_candidates": 3,
+        "storage_bpw": 16.0
       },
       "tensor": "model.language_model.layers.39.mlp.shared_expert_gate.weight"
     },
@@ -30343,8 +30343,8 @@
     "minimum_mtp_speedup": 1.2,
     "minimum_quality_retention": 0.98
   },
-  "created_at": "2026-08-02T04:42:36.690943Z",
-  "effective_bpw": 5.999999988428959,
+  "created_at": "2026-08-05T13:38:03.627894Z",
+  "effective_bpw": 5.999999959946398,
   "evidence_kind": "architecture_prior",
   "global_validation_required": true,
   "group_size": 32,
@@ -30368,6 +30368,7 @@
       "affine",
       "awq",
       "dwq",
+      "gptq",
       "bf16"
     ]
   },
@@ -30390,7 +30391,7 @@
       "parameters": 844640768
     }
   },
-  "nominal_bpw": 5.166136857126174,
+  "nominal_bpw": 5.166135034242241,
   "objective": {
     "cosine_distance": 0.05,
     "decode_latency_cost": 0.05,
@@ -30410,7 +30411,7 @@
   "schema_version": "axquant.plan.v1",
   "software_versions": {
     "ax_engine": null,
-    "axquant": "1.0.0",
+    "axquant": "1.2.0",
     "mlx": "0.32.0",
     "mlx_lm": "0.31.3",
     "pydantic": "2.13.4",
@@ -30426,7 +30427,7 @@
   },
   "status": "planned",
   "target_bpw": 6.0,
-  "target_class": "4bit",
+  "target_class": "6bit",
   "target_mode": "low-memory",
   "warnings": [
     "This report contains architecture priors, not calibration measurements.",
@@ -30441,16 +30442,16 @@
       "parameters": 24696061952
     },
     "6bit": {
-      "fraction": 0.2434611881590681,
-      "parameters": 8752873472
+      "fraction": 0.2434625553220185,
+      "parameters": 8752922624
     },
     "8bit": {
-      "fraction": 0.019524909815543244,
-      "parameters": 701956096
+      "fraction": 0.019523428722346983,
+      "parameters": 701902848
     },
     "bf16": {
-      "fraction": 0.05009290346215543,
-      "parameters": 1800931184
+      "fraction": 0.0500930173924013,
+      "parameters": 1800935280
     }
   }
 }