{
  "format_version": 1,
  "tool_version": "1.1",
  "model_path": "/home/michal/bench/models/Qwen_Qwen3.5-9B-Q4_K_M.gguf",
  "model_file_size": 6169341984,
  "fingerprints": {
    "model": {
      "algorithm": "fnv1a64",
      "scope": "first_1048576_bytes",
      "value": "f57a9ded6b417729"
    },
    "corpus": {
      "algorithm": "fnv1a64",
      "scope": "full_content",
      "value": "cce0b072d0f935e4"
    }
  },
  "architecture": "qwen35",
  "decoder_layer_count": 33,
  "corpus_source": "/home/michal/bench/corpus-calib.txt",
  "scoring_setup": {
    "chunk_size": 256,
    "screening_chunks": 6,
    "calibration_chunks": 17,
    "held_out_chunks": 17,
    "metric": "teacher-forced next-token NLL",
    "quality_budget_nats": 0.050000000000000003
  },
  "global_screening_candidate_count": 8,
  "global_shortlist_size": 8,
  "global_shortlist_units": [
    {
      "name": "gated-attention layer 15",
      "exact_layers": [15],
      "screening_nll": 2.6267523275280169,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 11",
      "exact_layers": [11],
      "screening_nll": 2.6340010706842238,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 7",
      "exact_layers": [7],
      "screening_nll": 2.6434640468359083,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 23",
      "exact_layers": [23],
      "screening_nll": 2.6926314663944995,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 27",
      "exact_layers": [27],
      "screening_nll": 2.6936436150725029,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 19",
      "exact_layers": [19],
      "screening_nll": 2.695413764782181,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 3",
      "exact_layers": [3],
      "screening_nll": 2.785657808286643,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "gated-attention layer 31",
      "exact_layers": [31],
      "screening_nll": 2.9274145530557583,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    }
  ],
  "baseline_screening_nll": 2.606246265026364,
  "baseline_calibration_nll": 2.6990915143454792,
  "baseline_held_out_nll": 2.4941308225298329,
  "protected_layers": [
    {
      "layer": 0,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 1,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 2,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 4,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 5,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 6,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 8,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 9,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 10,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 12,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 13,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 14,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 16,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 17,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 18,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 20,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 21,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 22,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 24,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 25,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 26,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 28,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 29,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 30,
      "capability": "stateful_non_kv",
      "reason": "recurrent state cannot be dropped"
    },
    {
      "layer": 32,
      "capability": "none",
      "reason": "layer is outside the supported droppable decoder-attention set"
    }
  ],
  "quality_cliff_heuristic": "Caution is flagged if the incremental delta NLL is > 0.005 and > 3x the median of previous positive increments (requires at least 3 previous positive steps). This detects a sudden cliff, not steady degradation - the 'recommended' profile is bounded by quality_budget_nats instead.",
  "steps": [
    {
      "step": 1,
      "drop_unit": "gated-attention layer 15",
      "exact_layers_dropped_in_step": [15],
      "cumulative_mask": "15",
      "incremental_delta_nll": 0.020869966693568553,
      "cumulative_delta_nll": 0.020869966693568553,
      "cumulative_dropped_attention_count": 1,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 7,
      "risk_status": "low estimated risk"
    },
    {
      "step": 2,
      "drop_unit": "gated-attention layer 23",
      "exact_layers_dropped_in_step": [23],
      "cumulative_mask": "15,23",
      "incremental_delta_nll": 0.051843655163706259,
      "cumulative_delta_nll": 0.072713621857274813,
      "cumulative_dropped_attention_count": 2,
      "cumulative_phys_kv_removed": 2,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 6,
      "risk_status": "low estimated risk"
    },
    {
      "step": 3,
      "drop_unit": "gated-attention layer 11",
      "exact_layers_dropped_in_step": [11],
      "cumulative_mask": "11,15,23",
      "incremental_delta_nll": 0.060156390229165524,
      "cumulative_delta_nll": 0.13287001208644034,
      "cumulative_dropped_attention_count": 3,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 5,
      "risk_status": "low estimated risk"
    },
    {
      "step": 4,
      "drop_unit": "gated-attention layer 7",
      "exact_layers_dropped_in_step": [7],
      "cumulative_mask": "7,11,15,23",
      "incremental_delta_nll": 0.1053080272093756,
      "cumulative_delta_nll": 0.23817803929581594,
      "cumulative_dropped_attention_count": 4,
      "cumulative_phys_kv_removed": 4,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 4,
      "risk_status": "low estimated risk"
    },
    {
      "step": 5,
      "drop_unit": "gated-attention layer 27",
      "exact_layers_dropped_in_step": [27],
      "cumulative_mask": "7,11,15,23,27",
      "incremental_delta_nll": 0.12857998022644868,
      "cumulative_delta_nll": 0.36675801952226461,
      "cumulative_dropped_attention_count": 5,
      "cumulative_phys_kv_removed": 5,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 3,
      "risk_status": "low estimated risk"
    },
    {
      "step": 6,
      "drop_unit": "gated-attention layer 19",
      "exact_layers_dropped_in_step": [19],
      "cumulative_mask": "7,11,15,19,23,27",
      "incremental_delta_nll": 0.23279800892556191,
      "cumulative_delta_nll": 0.59955602844782652,
      "cumulative_dropped_attention_count": 6,
      "cumulative_phys_kv_removed": 6,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 2,
      "risk_status": "caution"
    },
    {
      "step": 7,
      "drop_unit": "gated-attention layer 3",
      "exact_layers_dropped_in_step": [3],
      "cumulative_mask": "3,7,11,15,19,23,27",
      "incremental_delta_nll": 0.72605735665073245,
      "cumulative_delta_nll": 1.325613385098559,
      "cumulative_dropped_attention_count": 7,
      "cumulative_phys_kv_removed": 7,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 1,
      "risk_status": "caution"
    }
  ],
  "profiles": {
    "minimal": {
      "step": 1,
      "mask": "15",
      "held_out_delta_nll": 0.020869966693568553,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_1"
    },
    "light": {
      "step": 3,
      "mask": "11,15,23",
      "held_out_delta_nll": 0.13287001208644034,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_3"
    },
    "recommended": {
      "step": 1,
      "mask": "15",
      "held_out_delta_nll": 0.020869966693568553,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "within_quality_budget"
    },
    "max": {
      "step": 7,
      "mask": "3,7,11,15,19,23,27",
      "held_out_delta_nll": 1.325613385098559,
      "cumulative_phys_kv_removed": 7,
      "cumulative_shared_kv_removed": 0,
      "anchor": "deepest_legal_step",
      "warning": "deepest legal schedule; quality risk"
    }
  },
  "architecture_constraints": [
    {
      "kind": "qwen35_minimum_active_gated_attention",
      "required_active_layers": 1,
      "final_active_gated_attention_layers": 1,
      "reason": "the current hybrid memory path requires active attention K/V"
    }
  ]
}