{
  "format_version": 1,
  "tool_version": "1.1",
  "model_path": "/home/michal/bench/models/Llama-3.2-3B-Instruct-Q4_K_M.gguf",
  "model_file_size": 2019377696,
  "fingerprints": {
    "model": {
      "algorithm": "fnv1a64",
      "scope": "first_1048576_bytes",
      "value": "054f4b97ba816cae"
    },
    "corpus": {
      "algorithm": "fnv1a64",
      "scope": "full_content",
      "value": "cce0b072d0f935e4"
    }
  },
  "architecture": "llama",
  "decoder_layer_count": 28,
  "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": 28,
  "global_shortlist_size": 12,
  "global_shortlist_units": [
    {
      "name": "layer 8",
      "exact_layers": [8],
      "screening_nll": 2.8093987715419071,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 22",
      "exact_layers": [22],
      "screening_nll": 2.8145014459127755,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 21",
      "exact_layers": [21],
      "screening_nll": 2.8180902006889106,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 20",
      "exact_layers": [20],
      "screening_nll": 2.8247655987896967,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 7",
      "exact_layers": [7],
      "screening_nll": 2.8320837635575544,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 11",
      "exact_layers": [11],
      "screening_nll": 2.838483105558975,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 18",
      "exact_layers": [18],
      "screening_nll": 2.8387298884058159,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 10",
      "exact_layers": [10],
      "screening_nll": 2.8398288486284353,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 6",
      "exact_layers": [6],
      "screening_nll": 2.8454148257139291,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 17",
      "exact_layers": [17],
      "screening_nll": 2.8462361832767051,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 23",
      "exact_layers": [23],
      "screening_nll": 2.8472285550141621,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 19",
      "exact_layers": [19],
      "screening_nll": 2.8478483332872355,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    }
  ],
  "baseline_screening_nll": 2.8038607740276507,
  "baseline_calibration_nll": 3.1615423215248106,
  "baseline_held_out_nll": 2.8956674010049714,
  "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 8",
      "exact_layers_dropped_in_step": [8],
      "cumulative_mask": "8",
      "incremental_delta_nll": 0.0063824688837033072,
      "cumulative_delta_nll": 0.0063824688837033072,
      "cumulative_dropped_attention_count": 1,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 27,
      "risk_status": "low estimated risk"
    },
    {
      "step": 2,
      "drop_unit": "layer 21",
      "exact_layers_dropped_in_step": [21],
      "cumulative_mask": "8,21",
      "incremental_delta_nll": 0.014018064697843613,
      "cumulative_delta_nll": 0.02040053358154692,
      "cumulative_dropped_attention_count": 2,
      "cumulative_phys_kv_removed": 2,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 26,
      "risk_status": "low estimated risk"
    },
    {
      "step": 3,
      "drop_unit": "layer 18",
      "exact_layers_dropped_in_step": [18],
      "cumulative_mask": "8,18,21",
      "incremental_delta_nll": 0.01486791315251601,
      "cumulative_delta_nll": 0.03526844673406293,
      "cumulative_dropped_attention_count": 3,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 25,
      "risk_status": "low estimated risk"
    },
    {
      "step": 4,
      "drop_unit": "layer 19",
      "exact_layers_dropped_in_step": [19],
      "cumulative_mask": "8,18,19,21",
      "incremental_delta_nll": 0.053608182883182476,
      "cumulative_delta_nll": 0.088876629617245406,
      "cumulative_dropped_attention_count": 4,
      "cumulative_phys_kv_removed": 4,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 24,
      "risk_status": "caution"
    },
    {
      "step": 5,
      "drop_unit": "layer 23",
      "exact_layers_dropped_in_step": [23],
      "cumulative_mask": "8,18,19,21,23",
      "incremental_delta_nll": 0.0588957752894399,
      "cumulative_delta_nll": 0.14777240490668531,
      "cumulative_dropped_attention_count": 5,
      "cumulative_phys_kv_removed": 5,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 23,
      "risk_status": "caution"
    },
    {
      "step": 6,
      "drop_unit": "layer 22",
      "exact_layers_dropped_in_step": [22],
      "cumulative_mask": "8,18,19,21,22,23",
      "incremental_delta_nll": 0.043234023010637834,
      "cumulative_delta_nll": 0.19100642791732314,
      "cumulative_dropped_attention_count": 6,
      "cumulative_phys_kv_removed": 6,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 22,
      "risk_status": "low estimated risk"
    },
    {
      "step": 7,
      "drop_unit": "layer 11",
      "exact_layers_dropped_in_step": [11],
      "cumulative_mask": "8,11,18,19,21,22,23",
      "incremental_delta_nll": 0.034177364948544398,
      "cumulative_delta_nll": 0.22518379286586754,
      "cumulative_dropped_attention_count": 7,
      "cumulative_phys_kv_removed": 7,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 21,
      "risk_status": "low estimated risk"
    },
    {
      "step": 8,
      "drop_unit": "layer 17",
      "exact_layers_dropped_in_step": [17],
      "cumulative_mask": "8,11,17,18,19,21,22,23",
      "incremental_delta_nll": 0.072754178947235459,
      "cumulative_delta_nll": 0.297937971813103,
      "cumulative_dropped_attention_count": 8,
      "cumulative_phys_kv_removed": 8,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 20,
      "risk_status": "low estimated risk"
    },
    {
      "step": 9,
      "drop_unit": "layer 20",
      "exact_layers_dropped_in_step": [20],
      "cumulative_mask": "8,11,17,18,19,20,21,22,23",
      "incremental_delta_nll": 0.10413032905601938,
      "cumulative_delta_nll": 0.40206830086912237,
      "cumulative_dropped_attention_count": 9,
      "cumulative_phys_kv_removed": 9,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 19,
      "risk_status": "low estimated risk"
    },
    {
      "step": 10,
      "drop_unit": "layer 6",
      "exact_layers_dropped_in_step": [6],
      "cumulative_mask": "6,8,11,17,18,19,20,21,22,23",
      "incremental_delta_nll": 0.10544059337673684,
      "cumulative_delta_nll": 0.50750889424585921,
      "cumulative_dropped_attention_count": 10,
      "cumulative_phys_kv_removed": 10,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 18,
      "risk_status": "low estimated risk"
    },
    {
      "step": 11,
      "drop_unit": "layer 10",
      "exact_layers_dropped_in_step": [10],
      "cumulative_mask": "6,8,10,11,17,18,19,20,21,22,23",
      "incremental_delta_nll": 0.2310971113477156,
      "cumulative_delta_nll": 0.73860600559357481,
      "cumulative_dropped_attention_count": 11,
      "cumulative_phys_kv_removed": 11,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 17,
      "risk_status": "caution"
    },
    {
      "step": 12,
      "drop_unit": "layer 7",
      "exact_layers_dropped_in_step": [7],
      "cumulative_mask": "6,7,8,10,11,17,18,19,20,21,22,23",
      "incremental_delta_nll": 1.0073771489534282,
      "cumulative_delta_nll": 1.745983154547003,
      "cumulative_dropped_attention_count": 12,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 16,
      "risk_status": "caution"
    }
  ],
  "profiles": {
    "minimal": {
      "step": 1,
      "mask": "8",
      "held_out_delta_nll": 0.0063824688837033072,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_1"
    },
    "light": {
      "step": 3,
      "mask": "8,18,21",
      "held_out_delta_nll": 0.03526844673406293,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_3"
    },
    "recommended": {
      "step": 3,
      "mask": "8,18,21",
      "held_out_delta_nll": 0.03526844673406293,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "anchor": "within_quality_budget"
    },
    "max": {
      "step": 12,
      "mask": "6,7,8,10,11,17,18,19,20,21,22,23",
      "held_out_delta_nll": 1.745983154547003,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "anchor": "deepest_legal_step",
      "warning": "deepest legal schedule; quality risk"
    }
  }
}