{
  "version": "1.1",
  "strategy": "greedy",
  "output_tokens": 128,
  "prefix_max_tokens": 16,
  "inputs_per_dataset": 5,
  "repeats": 3,
  "datasets": [
    "GK",
    "WT",
    "HE",
    "GP",
    "C4"
  ],
  "workload_basis": "25 inputs, three repeats; historical input selection is not fully recovered.",
  "decoding_implementation": "Current corrected greedy argmax; existing native-precision normalization retained.",
  "primary_metric": "(last_commit_ms-first_commit_ms)/(N-1)",
  "additional_metrics": [
    "post_first_batch_ms_per_token",
    "ttft_ms",
    "e2e_ms_per_token"
  ],
  "cache_reset": "initial resident set and LRU restored before each input outside measurement",
  "numerical_determinism": "Native precision; target checks are recorded. Use --require-greedy-match to enforce equality.",
  "hardware": "NVIDIA A100 40GB; sequential cases; optional host placement check via --check-numa",
  "default_case": "ds_high",
  "supporting_sweeps": {
    "scope": "Reduced supporting experiments; not a reconstruction of every historical paper sweep.",
    "dataset": "GK",
    "inputs": 3,
    "selection": "First three configured GK inputs; depth reuses main/inputs.json when available.",
    "repeats": 3,
    "output_tokens": 128,
    "model": "dsv2lite",
    "offload_per_layer": 48,
    "depth": {
      "candidates": [2, 4, 6, 8, 12, 16, 20, 24, 32],
      "prefix_max_tokens": 16,
      "expected_records": 81
    },
    "prefix": {
      "prefix_max_tokens": [8, 16, 32],
      "gamma": 16,
      "expected_records_per_prefix": 9,
      "expected_records_total": 27
    }
  },
  "reference_budget_is_enforced_process_cap": false,
  "ds_low_selection": "Resident=4; ds_low8 remains available as an explicit alternative.",
  "cases": {
    "ds_high": {
      "status": "ready",
      "role": "main",
      "model": "dsv2lite",
      "tier": "high",
      "reference_budget_gb": 24,
      "gamma": 16,
      "resident_per_layer": 16,
      "offload_per_layer": 48,
      "buffer_size": 32
    },
    "ds_low4": {
      "status": "ready",
      "role": "main",
      "model": "dsv2lite",
      "tier": "low",
      "reference_budget_gb": 18,
      "gamma": 16,
      "resident_per_layer": 4,
      "offload_per_layer": 60,
      "buffer_size": 32
    },
    "ds_low8": {
      "status": "ready",
      "role": "main",
      "model": "dsv2lite",
      "tier": "low",
      "reference_budget_gb": 18,
      "gamma": 16,
      "resident_per_layer": 8,
      "offload_per_layer": 56,
      "buffer_size": 32,
      "original_paper_configuration_confirmed": false
    },
    "qwen_high": {
      "status": "ready",
      "role": "main",
      "model": "qwen2moe",
      "tier": "high",
      "reference_budget_gb": 20,
      "gamma": 8,
      "resident_per_layer": 15,
      "offload_per_layer": 45,
      "buffer_size": 32
    },
    "qwen_low": {
      "status": "ready",
      "role": "main",
      "model": "qwen2moe",
      "tier": "low",
      "reference_budget_gb": 16,
      "gamma": 8,
      "resident_per_layer": 8,
      "offload_per_layer": 52,
      "buffer_size": 32
    },
    "phi_high": {
      "status": "ready",
      "role": "main",
      "model": "phimoe",
      "tier": "high",
      "reference_budget_gb": 40,
      "gamma": 16,
      "resident_per_layer": 2,
      "offload_per_layer": 14,
      "buffer_size": 0
    },
    "phi_low": {
      "status": "ready",
      "role": "main",
      "model": "phimoe",
      "tier": "low",
      "reference_budget_gb": 32,
      "gamma": 16,
      "resident_per_layer": 1,
      "offload_per_layer": 15,
      "buffer_size": 0
    }
  }
}