{
  "schema_version": "edgeflowertune.benchmark_release.v1",
  "release_id": "edgeflowertune-20261020-v1",
  "status": "organizer_approved",
  "approved_on": "2026-10-20",
  "approval_scope": "Use the verified existing model and dataset bytes, splits, preprocessing and calibration for new submissions. This does not certify every historical result or a training convergence claim.",
  "asset_catalog": {
    "path": "participant_toolkit/reference_assets.v1.json",
    "url": "/downloads/reference-assets.json",
    "sha256": "5e87580f5a0720c6e609750ee11f1a993a56e38766bba6bb64687b80c4d9ef9e"
  },
  "protocol_policy": {
    "path": "webapp/data/official_protocol_policies.v1.json",
    "url": "/api/submissions/protocol-policy",
    "sha256": "30dbfef5e397bd9710c6a7f66938467f96ecb130977d09b8084d3774d96b45a8",
    "protocol_b_policy_version": "protocol-b-paper-tables-20261022-v1"
  },
  "models": [
    "qwen25_05b",
    "gemma3_270m",
    "gemma3_1b"
  ],
  "datasets": [
    "boolq",
    "qnli",
    "piqa",
    "hellaswag",
    "socialiqa",
    "arce",
    "winogrande"
  ],
  "data_contract": {
    "training_split": "train",
    "evaluation_split": "validation",
    "source_identity": "Original ordered CSV bytes; all fourteen hashes are in the asset catalog.",
    "quality_sample_id": "<dataset>/validation/<zero-based-source-row-index>",
    "client_partition": "raw source row index modulo client count, before tokenizer filtering",
    "sequence_length": {
      "boolq": 128,
      "other_registered_datasets": 64
    },
    "effective_batch_size": 8,
    "training_preprocessing": "Retain archived BoolQ passage shortening and skip overlength training rows, with retained/skipped row indices recorded.",
    "evaluation_preprocessing": "Keep every validation row; right-truncate overlength prompts using the fixed profile; record clipped row indices.",
    "explicit_run_settings": [
      "clients",
      "seed",
      "rounds",
      "local_steps"
    ],
    "details": "integrations/pytorch/REFERENCE_MATRIX.md"
  },
  "ranking": {
    "system_costs": "calibrated_simulation",
    "centralized_protocols": [
      "A",
      "B"
    ],
    "other_builtin_method_protocols": [
      "A",
      "B",
      "C"
    ],
    "historical_b_costs": "Paper table targets and their original costs participate in current Protocol B scoring."
  }
}