{
  "analysis_sha256": "c263d20cb4e55f11624c0ffd8f2b61210e3a86bd761284c50558c6367e710d12",
  "architecture_profile": {
    "adapter_id": "gpt-oss-v1",
    "audio_present": false,
    "config_model_type": "gpt_oss",
    "dense": false,
    "mtp_declared": false,
    "notes": [
      "OpenAI GPT-OSS MoE (model_type=gpt_oss): fused experts via MLX-LM SwitchGLU.",
      "Native MXFP4 expert packs convert with --allow-quantized (dequant → affine).",
      "No declared MTP; development evidence until a revision is certified.",
      "MoE experts quantize as fused switch modules with uniform per-group precision; artifacts are development evidence until certified."
    ],
    "optimization_scope": "text-path",
    "product_family": "gpt-oss",
    "support_level": "supported",
    "support_tier": "convertible",
    "text_layer_count": 24,
    "vision_present": false
  },
  "assignments": [
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.0.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.0.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.0.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.0.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.0.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.0.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.0.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.0.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.0.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.1.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.1.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.1.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.1.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.1.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.1.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.1.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.1.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.1.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.10.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.10.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.10.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.10.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.10.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.10.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.10.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.10.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.10.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.11.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.11.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.11.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.11.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.11.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.11.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.11.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.11.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.11.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.12.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.12.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.12.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.12.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.12.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.12.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.12.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.12.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.12.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.13.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.13.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.13.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.13.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.13.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.13.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.13.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.13.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.13.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.14.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.14.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.14.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.14.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.14.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.14.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.14.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.14.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.14.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.15.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.15.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.15.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.15.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.15.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.15.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.15.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.15.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.15.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.16.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.16.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.16.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.16.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.16.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.16.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.16.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.16.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.16.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.17.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.17.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.17.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.17.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.17.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.17.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.17.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.17.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.17.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.self_attn.k_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.18.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.self_attn.q_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.self_attn.v_proj.bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.18.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.18.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.18.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.post_attention_layernorm.weight"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.18.self_attn.k_proj.weight"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.18.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.18.self_attn.sinks"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.18.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.18.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.19.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.19.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.19.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.19.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.19.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.19.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.19.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.19.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.19.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.2.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.2.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.2.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.2.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.2.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.2.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.2.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.2.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.2.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.20.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.20.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.20.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.20.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.20.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.20.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.20.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.20.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.20.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.21.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.21.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.21.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.21.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.21.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.21.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.21.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.21.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.21.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.22.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.22.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.22.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.22.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.22.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.22.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.22.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.22.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.22.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.23.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.23.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.23.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.23.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.23.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.23.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.23.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.23.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.23.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.3.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.3.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.3.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.3.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.3.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.3.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.3.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.3.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.3.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.4.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.4.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.4.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.4.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.4.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.4.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.4.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.4.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.4.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.5.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.5.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.5.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.5.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.5.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.5.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.5.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.5.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.5.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.6.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.6.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.6.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.6.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.6.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "lm_head",
      "outlier_strategy": "none",
      "parameters": 579133440,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-lm-head: Protection floor — LM head stays BF16",
      "role": "lm_head",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "lm_head.weight"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.embed_tokens",
      "outlier_strategy": "none",
      "parameters": 579133440,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-embeddings: Embedding floor 8-bit",
      "role": "embedding",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.embed_tokens.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.6.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.6.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.6.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.6.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.7.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.7.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.7.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.7.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.7.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.7.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.7.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.7.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.7.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.8.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.8.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.8.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.8.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.8.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.8.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.8.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.8.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.8.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.input_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.input_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.mlp.experts.down_proj_bias",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.mlp.experts.down_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.mlp.experts.down_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 265420800,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.9.mlp.experts.down_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.mlp.experts.gate_up_proj_bias",
      "outlier_strategy": "none",
      "parameters": 184320,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "expert",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.mlp.experts.gate_up_proj_bias"
    },
    {
      "bits": 4,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.mlp.experts.gate_up_proj_blocks",
      "outlier_strategy": "none",
      "parameters": 530841600,
      "predicted_loss": 0.0,
      "reason": "manual rule experts-4bit: MoE experts/MLP at 4-bit product trunk",
      "role": "expert",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 4.5
      },
      "tensor": "model.layers.9.mlp.experts.gate_up_proj_blocks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.mlp.router.bias",
      "outlier_strategy": "none",
      "parameters": 32,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "router",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.mlp.router.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.mlp.router",
      "outlier_strategy": "none",
      "parameters": 92160,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-routers: Router floor 8-bit for routing stability",
      "role": "router",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.9.mlp.router.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.post_attention_layernorm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.post_attention_layernorm.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.k_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.self_attn.k_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.k_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.9.self_attn.k_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.o_proj.bias",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.self_attn.o_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.o_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.9.self_attn.o_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.q_proj.bias",
      "outlier_strategy": "none",
      "parameters": 4096,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.self_attn.q_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.q_proj",
      "outlier_strategy": "none",
      "parameters": 11796480,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.9.self_attn.q_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.sinks",
      "outlier_strategy": "none",
      "parameters": 64,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.self_attn.sinks"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.v_proj.bias",
      "outlier_strategy": "none",
      "parameters": 512,
      "predicted_loss": 0.0,
      "reason": "non-quantizable tensor preserved",
      "role": "attention",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.layers.9.self_attn.v_proj.bias"
    },
    {
      "bits": 8,
      "group_size": 64,
      "method": "affine",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.layers.9.self_attn.v_proj",
      "outlier_strategy": "none",
      "parameters": 1474560,
      "predicted_loss": 0.0,
      "reason": "manual rule attention-8bit: Raise attention for general + coding fidelity",
      "role": "attention",
      "scale_strategy": "group-affine",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 8.5
      },
      "tensor": "model.layers.9.self_attn.v_proj.weight"
    },
    {
      "bits": 16,
      "group_size": null,
      "method": "bf16",
      "metrics": {
        "cosine_distance": 0.0,
        "decode_latency_cost": 0.0,
        "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": 0.0,
        "token_disagreement": 0.0
      },
      "module_path": "model.norm",
      "outlier_strategy": "none",
      "parameters": 2880,
      "predicted_loss": 0.0,
      "reason": "manual rule protect-norms: Protection floor — norms stay BF16",
      "role": "norm",
      "scale_strategy": "none",
      "strategy_metadata": {
        "selected_from_candidates": 1,
        "storage_bpw": 16.0
      },
      "tensor": "model.norm.weight"
    }
  ],
  "calibration": null,
  "candidate_bits": [
    4,
    8,
    16
  ],
  "candidate_group_sizes": [],
  "constraints": {
    "effective_bpw_limit": 5.2,
    "lm_head_min_bits": 16,
    "max_model_size_ratio_to_uniform4": 1.1,
    "minimum_mtp_acceptance_retention": 0.95,
    "minimum_mtp_speedup": 1.2,
    "minimum_quality_retention": 0.98
  },
  "cost_model": "abstract-bpw",
  "created_at": "2026-08-11T22:44:24.531951Z",
  "effective_bpw": 5.055283578663038,
  "evidence_kind": "architecture_prior",
  "global_validation_required": true,
  "group_size": 64,
  "hardware": {
    "name": "ax-engine-apple-silicon-affine-dwq-v3",
    "runtime": "ax-engine",
    "supported_bits": [
      2,
      3,
      4,
      6,
      8,
      16
    ],
    "supported_group_sizes": [
      32,
      64,
      128
    ],
    "supported_methods": [
      "affine",
      "awq",
      "dwq",
      "gptq",
      "gptq-act",
      "bf16"
    ]
  },
  "kernel_latency_host_id": null,
  "kernel_latency_sha256": null,
  "kv_cache": null,
  "method_near_ties": [],
  "method_near_ties_omitted": 0,
  "mtp": {
    "candidate_bits": [
      8,
      16
    ],
    "min_bits": 8,
    "mode": "protected",
    "optimize_for_acceptance": true,
    "preserve_external_sidecar": true,
    "protect_norms": true,
    "protect_output_head": true
  },
  "mtp_distribution": {},
  "nominal_bpw": 4.5692953211576715,
  "objective": {
    "cosine_distance": 0.03,
    "decode_latency_cost": 0.07,
    "hidden_state_error": 0.1,
    "long_context_loss": 0.05,
    "mtp_acceptance_loss": 0.22,
    "output_kl": 0.15,
    "peak_memory_cost": 0.03,
    "prefill_latency_cost": 0.03,
    "task_loss_delta": 0.2,
    "token_disagreement": 0.12
  },
  "primary_runtime": "ax-engine",
  "profile": "agent-coding",
  "quantizer": "axquant",
  "random_seed": 20260728,
  "schema_version": "axquant.plan.v1",
  "software_versions": {
    "ax_engine": "6.11.1",
    "axquant": "1.6.2",
    "mlx": "0.32.0",
    "mlx_lm": "0.31.3",
    "pydantic": "2.13.4",
    "python": "3.12.13",
    "safetensors": "0.8.0"
  },
  "source_model": {
    "architecture": "GptOssForCausalLM",
    "format": "mlx",
    "local_path": null,
    "model_id": "openai/gpt-oss-20b",
    "revision": "6cee5e81ee83917806bbde320786a8fb61efebee"
  },
  "status": "planned",
  "target_bpw": 5.2,
  "target_class": "4bit",
  "target_mode": "balanced",
  "warnings": [
    "Manual assignments are unmeasured development evidence.",
    "Conversion requires --allow-unmeasured and cannot pass publication gates."
  ],
  "weight_distribution": {
    "4bit": {
      "fraction": 0.9137231396891172,
      "parameters": 19110297600
    },
    "8bit": {
      "fraction": 0.05825337532161521,
      "parameters": 1218355200
    },
    "bf16": {
      "fraction": 0.028023484989267565,
      "parameters": 586104384
    }
  }
}