{
  "format_version": 1,
  "tool_version": "1.1",
  "model_path": "/home/michal/bench/models/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
  "model_file_size": 4920739232,
  "fingerprints": {
    "model": {
      "algorithm": "fnv1a64",
      "scope": "first_1048576_bytes",
      "value": "5bbb84624e594409"
    },
    "corpus": {
      "algorithm": "fnv1a64",
      "scope": "full_content",
      "value": "cce0b072d0f935e4"
    }
  },
  "architecture": "llama",
  "decoder_layer_count": 32,
  "corpus_source": "/home/michal/bench/corpus-calib.txt",
  "scoring_setup": {
    "chunk_size": 256,
    "screening_chunks": 6,
    "calibration_chunks": 16,
    "held_out_chunks": 17,
    "metric": "teacher-forced next-token NLL",
    "quality_budget_nats": 0.050000000000000003
  },
  "global_screening_candidate_count": 32,
  "global_shortlist_size": 12,
  "global_shortlist_units": [
    {
      "name": "layer 24",
      "exact_layers": [24],
      "screening_nll": 2.4754291604138081,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 23",
      "exact_layers": [23],
      "screening_nll": 2.4790538703942757,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 25",
      "exact_layers": [25],
      "screening_nll": 2.480538205141595,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 22",
      "exact_layers": [22],
      "screening_nll": 2.4816702249450158,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 28",
      "exact_layers": [28],
      "screening_nll": 2.4836992149544357,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 21",
      "exact_layers": [21],
      "screening_nll": 2.4836772133058198,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 20",
      "exact_layers": [20],
      "screening_nll": 2.4881451415402585,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 26",
      "exact_layers": [26],
      "screening_nll": 2.4906905190959154,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 10",
      "exact_layers": [10],
      "screening_nll": 2.4923445501289549,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 19",
      "exact_layers": [19],
      "screening_nll": 2.4943386755979109,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 30",
      "exact_layers": [30],
      "screening_nll": 2.4952753891230341,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 27",
      "exact_layers": [27],
      "screening_nll": 2.4965655528440789,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    }
  ],
  "baseline_screening_nll": 2.4679272395671199,
  "baseline_calibration_nll": 2.8528067352768556,
  "baseline_held_out_nll": 2.5307234996686754,
  "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": 0.0048779675768440001,
      "cumulative_delta_nll": 0.0048779675768440001,
      "cumulative_dropped_attention_count": 1,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 31,
      "risk_status": "low estimated risk"
    },
    {
      "step": 2,
      "drop_unit": "layer 20",
      "exact_layers_dropped_in_step": [20],
      "cumulative_mask": "20,24",
      "incremental_delta_nll": 0.010340488471927589,
      "cumulative_delta_nll": 0.015218456048771589,
      "cumulative_dropped_attention_count": 2,
      "cumulative_phys_kv_removed": 2,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 30,
      "risk_status": "low estimated risk"
    },
    {
      "step": 3,
      "drop_unit": "layer 25",
      "exact_layers_dropped_in_step": [25],
      "cumulative_mask": "20,24,25",
      "incremental_delta_nll": 0.011283437562755072,
      "cumulative_delta_nll": 0.026501893611526661,
      "cumulative_dropped_attention_count": 3,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 29,
      "risk_status": "low estimated risk"
    },
    {
      "step": 4,
      "drop_unit": "layer 19",
      "exact_layers_dropped_in_step": [19],
      "cumulative_mask": "19,20,24,25",
      "incremental_delta_nll": 0.021414728248678738,
      "cumulative_delta_nll": 0.047916621860205399,
      "cumulative_dropped_attention_count": 4,
      "cumulative_phys_kv_removed": 4,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 28,
      "risk_status": "low estimated risk"
    },
    {
      "step": 5,
      "drop_unit": "layer 28",
      "exact_layers_dropped_in_step": [28],
      "cumulative_mask": "19,20,24,25,28",
      "incremental_delta_nll": 0.027529377024450685,
      "cumulative_delta_nll": 0.075445998884656085,
      "cumulative_dropped_attention_count": 5,
      "cumulative_phys_kv_removed": 5,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 27,
      "risk_status": "low estimated risk"
    },
    {
      "step": 6,
      "drop_unit": "layer 22",
      "exact_layers_dropped_in_step": [22],
      "cumulative_mask": "19,20,22,24,25,28",
      "incremental_delta_nll": 0.022608930309032882,
      "cumulative_delta_nll": 0.098054929193688967,
      "cumulative_dropped_attention_count": 6,
      "cumulative_phys_kv_removed": 6,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 26,
      "risk_status": "low estimated risk"
    },
    {
      "step": 7,
      "drop_unit": "layer 23",
      "exact_layers_dropped_in_step": [23],
      "cumulative_mask": "19,20,22,23,24,25,28",
      "incremental_delta_nll": 0.021069162379590001,
      "cumulative_delta_nll": 0.11912409157327897,
      "cumulative_dropped_attention_count": 7,
      "cumulative_phys_kv_removed": 7,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 25,
      "risk_status": "low estimated risk"
    },
    {
      "step": 8,
      "drop_unit": "layer 30",
      "exact_layers_dropped_in_step": [30],
      "cumulative_mask": "19,20,22,23,24,25,28,30",
      "incremental_delta_nll": 0.036364759452527817,
      "cumulative_delta_nll": 0.15548885102580678,
      "cumulative_dropped_attention_count": 8,
      "cumulative_phys_kv_removed": 8,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 24,
      "risk_status": "low estimated risk"
    },
    {
      "step": 9,
      "drop_unit": "layer 10",
      "exact_layers_dropped_in_step": [10],
      "cumulative_mask": "10,19,20,22,23,24,25,28,30",
      "incremental_delta_nll": 0.038926564709826028,
      "cumulative_delta_nll": 0.19441541573563281,
      "cumulative_dropped_attention_count": 9,
      "cumulative_phys_kv_removed": 9,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 23,
      "risk_status": "low estimated risk"
    },
    {
      "step": 10,
      "drop_unit": "layer 21",
      "exact_layers_dropped_in_step": [21],
      "cumulative_mask": "10,19,20,21,22,23,24,25,28,30",
      "incremental_delta_nll": 0.056603590704341844,
      "cumulative_delta_nll": 0.25101900643997466,
      "cumulative_dropped_attention_count": 10,
      "cumulative_phys_kv_removed": 10,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 22,
      "risk_status": "low estimated risk"
    },
    {
      "step": 11,
      "drop_unit": "layer 26",
      "exact_layers_dropped_in_step": [26],
      "cumulative_mask": "10,19,20,21,22,23,24,25,26,28,30",
      "incremental_delta_nll": 0.053638119633670822,
      "cumulative_delta_nll": 0.30465712607364548,
      "cumulative_dropped_attention_count": 11,
      "cumulative_phys_kv_removed": 11,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 21,
      "risk_status": "low estimated risk"
    },
    {
      "step": 12,
      "drop_unit": "layer 27",
      "exact_layers_dropped_in_step": [27],
      "cumulative_mask": "10,19,20,21,22,23,24,25,26,27,28,30",
      "incremental_delta_nll": 0.14117274356424048,
      "cumulative_delta_nll": 0.44582986963788596,
      "cumulative_dropped_attention_count": 12,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 20,
      "risk_status": "caution"
    }
  ],
  "profiles": {
    "minimal": {
      "step": 1,
      "mask": "24",
      "held_out_delta_nll": 0.0048779675768440001,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_1"
    },
    "light": {
      "step": 3,
      "mask": "20,24,25",
      "held_out_delta_nll": 0.026501893611526661,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_3"
    },
    "recommended": {
      "step": 4,
      "mask": "19,20,24,25",
      "held_out_delta_nll": 0.047916621860205399,
      "cumulative_phys_kv_removed": 4,
      "cumulative_shared_kv_removed": 0,
      "anchor": "within_quality_budget"
    },
    "max": {
      "step": 12,
      "mask": "10,19,20,21,22,23,24,25,26,27,28,30",
      "held_out_delta_nll": 0.44582986963788596,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "anchor": "deepest_legal_step",
      "warning": "deepest legal schedule; quality risk"
    }
  }
}