{
  "format_version": 1,
  "tool_version": "1.1",
  "model_path": "/home/michal/bench/models/Llama-3.2-1B-Instruct-Q4_K_M.gguf",
  "model_file_size": 807694464,
  "fingerprints": {
    "model": {
      "algorithm": "fnv1a64",
      "scope": "first_1048576_bytes",
      "value": "ab948931c1ff687e"
    },
    "corpus": {
      "algorithm": "fnv1a64",
      "scope": "full_content",
      "value": "cce0b072d0f935e4"
    }
  },
  "architecture": "llama",
  "decoder_layer_count": 16,
  "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": 16,
  "global_shortlist_size": 12,
  "global_shortlist_units": [
    {
      "name": "layer 13",
      "exact_layers": [13],
      "screening_nll": 3.223347689428175,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 10",
      "exact_layers": [10],
      "screening_nll": 3.2582729622720366,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 11",
      "exact_layers": [11],
      "screening_nll": 3.2736334592121712,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 8",
      "exact_layers": [8],
      "screening_nll": 3.2854423965377513,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 6",
      "exact_layers": [6],
      "screening_nll": 3.292278221236979,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 12",
      "exact_layers": [12],
      "screening_nll": 3.2984727643456218,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 14",
      "exact_layers": [14],
      "screening_nll": 3.3117470625626684,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 15",
      "exact_layers": [15],
      "screening_nll": 3.3116732687542965,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 5",
      "exact_layers": [5],
      "screening_nll": 3.3185929642141967,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 2",
      "exact_layers": [2],
      "screening_nll": 3.357962694709832,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 4",
      "exact_layers": [4],
      "screening_nll": 3.3678163709023141,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    },
    {
      "name": "layer 7",
      "exact_layers": [7],
      "screening_nll": 3.4143447950246255,
      "phys_kv_removed": 1,
      "shared_kv_removed": 0
    }
  ],
  "baseline_screening_nll": 3.1548324578014686,
  "baseline_calibration_nll": 3.4161869836291996,
  "baseline_held_out_nll": 3.1973892008368225,
  "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 13",
      "exact_layers_dropped_in_step": [13],
      "cumulative_mask": "13",
      "incremental_delta_nll": 0.080415718547751425,
      "cumulative_delta_nll": 0.080415718547751425,
      "cumulative_dropped_attention_count": 1,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 15,
      "risk_status": "low estimated risk"
    },
    {
      "step": 2,
      "drop_unit": "layer 10",
      "exact_layers_dropped_in_step": [10],
      "cumulative_mask": "10,13",
      "incremental_delta_nll": 0.11886333147055295,
      "cumulative_delta_nll": 0.19927905001830437,
      "cumulative_dropped_attention_count": 2,
      "cumulative_phys_kv_removed": 2,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 14,
      "risk_status": "low estimated risk"
    },
    {
      "step": 3,
      "drop_unit": "layer 14",
      "exact_layers_dropped_in_step": [14],
      "cumulative_mask": "10,13,14",
      "incremental_delta_nll": 0.17955455881501203,
      "cumulative_delta_nll": 0.3788336088333164,
      "cumulative_dropped_attention_count": 3,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 13,
      "risk_status": "low estimated risk"
    },
    {
      "step": 4,
      "drop_unit": "layer 8",
      "exact_layers_dropped_in_step": [8],
      "cumulative_mask": "8,10,13,14",
      "incremental_delta_nll": 0.1679762912090359,
      "cumulative_delta_nll": 0.5468099000423523,
      "cumulative_dropped_attention_count": 4,
      "cumulative_phys_kv_removed": 4,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 12,
      "risk_status": "low estimated risk"
    },
    {
      "step": 5,
      "drop_unit": "layer 6",
      "exact_layers_dropped_in_step": [6],
      "cumulative_mask": "6,8,10,13,14",
      "incremental_delta_nll": 0.22142265028979446,
      "cumulative_delta_nll": 0.76823255033214677,
      "cumulative_dropped_attention_count": 5,
      "cumulative_phys_kv_removed": 5,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 11,
      "risk_status": "low estimated risk"
    },
    {
      "step": 6,
      "drop_unit": "layer 11",
      "exact_layers_dropped_in_step": [11],
      "cumulative_mask": "6,8,10,11,13,14",
      "incremental_delta_nll": 0.2281282673898346,
      "cumulative_delta_nll": 0.99636081772198137,
      "cumulative_dropped_attention_count": 6,
      "cumulative_phys_kv_removed": 6,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 10,
      "risk_status": "low estimated risk"
    },
    {
      "step": 7,
      "drop_unit": "layer 12",
      "exact_layers_dropped_in_step": [12],
      "cumulative_mask": "6,8,10,11,12,13,14",
      "incremental_delta_nll": 0.41223330495740917,
      "cumulative_delta_nll": 1.4085941226793905,
      "cumulative_dropped_attention_count": 7,
      "cumulative_phys_kv_removed": 7,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 9,
      "risk_status": "low estimated risk"
    },
    {
      "step": 8,
      "drop_unit": "layer 2",
      "exact_layers_dropped_in_step": [2],
      "cumulative_mask": "2,6,8,10,11,12,13,14",
      "incremental_delta_nll": 0.52922927642554907,
      "cumulative_delta_nll": 1.9378233991049396,
      "cumulative_dropped_attention_count": 8,
      "cumulative_phys_kv_removed": 8,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 8,
      "risk_status": "low estimated risk"
    },
    {
      "step": 9,
      "drop_unit": "layer 15",
      "exact_layers_dropped_in_step": [15],
      "cumulative_mask": "2,6,8,10,11,12,13,14,15",
      "incremental_delta_nll": 0.59646051142911993,
      "cumulative_delta_nll": 2.5342839105340595,
      "cumulative_dropped_attention_count": 9,
      "cumulative_phys_kv_removed": 9,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 7,
      "risk_status": "low estimated risk"
    },
    {
      "step": 10,
      "drop_unit": "layer 7",
      "exact_layers_dropped_in_step": [7],
      "cumulative_mask": "2,6,7,8,10,11,12,13,14,15",
      "incremental_delta_nll": 0.85097357887000413,
      "cumulative_delta_nll": 3.3852574894040637,
      "cumulative_dropped_attention_count": 10,
      "cumulative_phys_kv_removed": 10,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 6,
      "risk_status": "caution"
    },
    {
      "step": 11,
      "drop_unit": "layer 4",
      "exact_layers_dropped_in_step": [4],
      "cumulative_mask": "2,4,6,7,8,10,11,12,13,14,15",
      "incremental_delta_nll": 0.34552198355373687,
      "cumulative_delta_nll": 3.7307794729578005,
      "cumulative_dropped_attention_count": 11,
      "cumulative_phys_kv_removed": 11,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 5,
      "risk_status": "low estimated risk"
    },
    {
      "step": 12,
      "drop_unit": "layer 5",
      "exact_layers_dropped_in_step": [5],
      "cumulative_mask": "2,4,5,6,7,8,10,11,12,13,14,15",
      "incremental_delta_nll": 0.61003634611002955,
      "cumulative_delta_nll": 4.3408158190678297,
      "cumulative_dropped_attention_count": 12,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "remaining_eligible_attention": 4,
      "risk_status": "low estimated risk"
    }
  ],
  "profiles": {
    "minimal": {
      "step": 1,
      "mask": "13",
      "held_out_delta_nll": 0.080415718547751425,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_1"
    },
    "light": {
      "step": 3,
      "mask": "10,13,14",
      "held_out_delta_nll": 0.3788336088333164,
      "cumulative_phys_kv_removed": 3,
      "cumulative_shared_kv_removed": 0,
      "anchor": "step_3"
    },
    "recommended": {
      "step": 1,
      "mask": "13",
      "held_out_delta_nll": 0.080415718547751425,
      "cumulative_phys_kv_removed": 1,
      "cumulative_shared_kv_removed": 0,
      "anchor": "budget_exceeded",
      "warning": "no schedule met the quality budget; showing the smallest drop unit"
    },
    "max": {
      "step": 12,
      "mask": "2,4,5,6,7,8,10,11,12,13,14,15",
      "held_out_delta_nll": 4.3408158190678297,
      "cumulative_phys_kv_removed": 12,
      "cumulative_shared_kv_removed": 0,
      "anchor": "deepest_legal_step",
      "warning": "deepest legal schedule; quality risk"
    }
  }
}