{
  "format_version": 1,
  "tool_version": "1.1",
  "model_path": "/home/michal/bench/models/google_gemma-3-12b-it-Q4_K_M.gguf",
  "model_file_size": 7300575264,
  "fingerprints": {
    "model": {
      "algorithm": "fnv1a64",
      "scope": "first_1048576_bytes",
      "value": "c5946a80b24a7cea"
    },
    "corpus": {
      "algorithm": "fnv1a64",
      "scope": "full_content",
      "value": "cce0b072d0f935e4"
    }
  },
  "architecture": "gemma3",
  "decoder_layer_count": 48,
  "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": 48,
  "global_shortlist_size": 12,
  "global_shortlist_units": [
    {
      "name": "layer 22",
      "exact_layers": [22],
      "screening_nll": 2.767811414692384,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 24",
      "exact_layers": [24],
      "screening_nll": 2.7713474515491043,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 17",
      "exact_layers": [17],
      "screening_nll": 2.8173819592852216,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 47",
      "exact_layers": [47],
      "screening_nll": 2.8207634239106096,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 26",
      "exact_layers": [26],
      "screening_nll": 2.8408762735741724,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 21",
      "exact_layers": [21],
      "screening_nll": 2.8416059895495169,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 45",
      "exact_layers": [45],
      "screening_nll": 2.847856316339247,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 12",
      "exact_layers": [12],
      "screening_nll": 2.8504535654164473,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 11",
      "exact_layers": [11],
      "screening_nll": 2.8529054157320703,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 6",
      "exact_layers": [6],
      "screening_nll": 2.8545234249565863,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 16",
      "exact_layers": [16],
      "screening_nll": 2.8559943032065482,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 1",
      "exact_layers": [1],
      "screening_nll": 2.8574715312199612,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    }
  ],
  "baseline_screening_nll": 2.8640220396141807,
  "baseline_calibration_nll": 3.0908448069924948,
  "baseline_held_out_nll": 2.798345287607213,
  "protected_layers": [

  ],
  "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": "layer 24",
      "exact_layers_dropped_in_step": [24],
      "cumulative_mask": "24",
      "incremental_delta_nll": -0.073397044894725649,
      "cumulative_delta_nll": -0.073397044894725649,
      "cumulative_dropped_attention_count": 1,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 47,
      "risk_status": "low estimated risk"
    },
    {
      "step": 2,
      "drop_unit": "layer 47",
      "exact_layers_dropped_in_step": [47],
      "cumulative_mask": "24,47",
      "incremental_delta_nll": -0.013307485262865182,
      "cumulative_delta_nll": -0.086704530157590831,
      "cumulative_dropped_attention_count": 2,
      "cumulative_phys_kv_removed": 2,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 46,
      "risk_status": "low estimated risk"
    },
    {
      "step": 3,
      "drop_unit": "layer 12",
      "exact_layers_dropped_in_step": [12],
      "cumulative_mask": "12,24,47",
      "incremental_delta_nll": 0.0026257713872297472,
      "cumulative_delta_nll": -0.084078758770361084,
      "cumulative_dropped_attention_count": 3,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 45,
      "risk_status": "low estimated risk"
    },
    {
      "step": 4,
      "drop_unit": "layer 1",
      "exact_layers_dropped_in_step": [1],
      "cumulative_mask": "1,12,24,47",
      "incremental_delta_nll": 0.03326274924966155,
      "cumulative_delta_nll": -0.050816009520699534,
      "cumulative_dropped_attention_count": 4,
      "cumulative_phys_kv_removed": 4,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 44,
      "risk_status": "low estimated risk"
    },
    {
      "step": 5,
      "drop_unit": "layer 45",
      "exact_layers_dropped_in_step": [45],
      "cumulative_mask": "1,12,24,45,47",
      "incremental_delta_nll": 0.030163028448804585,
      "cumulative_delta_nll": -0.020652981071894949,
      "cumulative_dropped_attention_count": 5,
      "cumulative_phys_kv_removed": 5,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 43,
      "risk_status": "low estimated risk"
    },
    {
      "step": 6,
      "drop_unit": "layer 16",
      "exact_layers_dropped_in_step": [16],
      "cumulative_mask": "1,12,16,24,45,47",
      "incremental_delta_nll": 0.010938357051516689,
      "cumulative_delta_nll": -0.0097146240203782597,
      "cumulative_dropped_attention_count": 6,
      "cumulative_phys_kv_removed": 6,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 42,
      "risk_status": "low estimated risk"
    },
    {
      "step": 7,
      "drop_unit": "layer 11",
      "exact_layers_dropped_in_step": [11],
      "cumulative_mask": "1,11,12,16,24,45,47",
      "incremental_delta_nll": 0.066409076208585827,
      "cumulative_delta_nll": 0.056694452188207567,
      "cumulative_dropped_attention_count": 7,
      "cumulative_phys_kv_removed": 7,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 41,
      "risk_status": "caution"
    },
    {
      "step": 8,
      "drop_unit": "layer 21",
      "exact_layers_dropped_in_step": [21],
      "cumulative_mask": "1,11,12,16,21,24,45,47",
      "incremental_delta_nll": 0.043000101254102585,
      "cumulative_delta_nll": 0.099694553442310152,
      "cumulative_dropped_attention_count": 8,
      "cumulative_phys_kv_removed": 8,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 40,
      "risk_status": "low estimated risk"
    },
    {
      "step": 9,
      "drop_unit": "layer 26",
      "exact_layers_dropped_in_step": [26],
      "cumulative_mask": "1,11,12,16,21,24,26,45,47",
      "incremental_delta_nll": 0.10056229775646974,
      "cumulative_delta_nll": 0.20025685119877989,
      "cumulative_dropped_attention_count": 9,
      "cumulative_phys_kv_removed": 9,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 39,
      "risk_status": "caution"
    },
    {
      "step": 10,
      "drop_unit": "layer 6",
      "exact_layers_dropped_in_step": [6],
      "cumulative_mask": "1,6,11,12,16,21,24,26,45,47",
      "incremental_delta_nll": 0.11911717903730867,
      "cumulative_delta_nll": 0.31937403023608857,
      "cumulative_dropped_attention_count": 10,
      "cumulative_phys_kv_removed": 10,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 38,
      "risk_status": "caution"
    },
    {
      "step": 11,
      "drop_unit": "layer 22",
      "exact_layers_dropped_in_step": [22],
      "cumulative_mask": "1,6,11,12,16,21,22,24,26,45,47",
      "incremental_delta_nll": 0.22538826170876458,
      "cumulative_delta_nll": 0.54476229194485315,
      "cumulative_dropped_attention_count": 11,
      "cumulative_phys_kv_removed": 11,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 37,
      "risk_status": "caution"
    },
    {
      "step": 12,
      "drop_unit": "layer 17",
      "exact_layers_dropped_in_step": [17],
      "cumulative_mask": "1,6,11,12,16,17,21,22,24,26,45,47",
      "incremental_delta_nll": 0.60635869513764185,
      "cumulative_delta_nll": 1.151120987082495,
      "cumulative_dropped_attention_count": 12,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 36,
      "risk_status": "caution"
    }
  ],
  "profiles": {
    "minimal": {
      "step": 1,
      "mask": "24",
      "held_out_delta_nll": -0.073397044894725649,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_1"
    },
    "light": {
      "step": 3,
      "mask": "12,24,47",
      "held_out_delta_nll": -0.084078758770361084,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_3"
    },
    "recommended": {
      "step": 6,
      "mask": "1,12,16,24,45,47",
      "held_out_delta_nll": -0.0097146240203782597,
      "cumulative_phys_kv_removed": 6,
      "cumulative_shared_kv_removed": 0,
      "anchor": "within_quality_budget"
    },
    "max": {
      "step": 12,
      "mask": "1,6,11,12,16,17,21,22,24,26,45,47",
      "held_out_delta_nll": 1.151120987082495,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "anchor": "deepest_legal_step",
      "warning": "deepest legal schedule; quality risk"
    }
  }
}