{
  "schema_version": 1,
  "adapter_key": "system-one-formal-evolve",
  "validated_at": "2026-09-26",
  "accelerator": {"vendor": "NVIDIA", "model": "A100-SXM4-80GB"},
  "command": [
    "python", "scripts/run_system_one_formal_evolve.py",
    "--data", "data/system-one/system-one-public-v1.jsonl",
    "--nimble", "checkpoint/nimble",
    "--nimble-base", "checkpoint/qwen3.5-9b",
    "--laya", "checkpoint/laya",
    "--generations", "2", "--population", "2", "--seed", "42", "--device", "cuda:0"
  ],
  "dataset": {
    "name": "System One public v1: BoolQ, HelpSteer2, PubMedQA, ARC-Challenge, TruthfulQA",
    "license": "CC BY-SA 3.0 / CC BY 4.0 / MIT / CC BY-SA 4.0 / Apache-2.0",
    "revision": "sha256:387638878933f8b6293b240a1daf5017fb317bd3c226e511c4e72cf3af46140d",
    "calibration_examples": 60,
    "test_examples": 60,
    "ood_examples": 80,
    "test_isolation": true
  },
  "checkpoint": {
    "models": [
      {"model_id": "bespokelabs/Bespoke-Nimble-9B", "revision": "594dfdcfb6f94e3d0c0db7535180d3c71689169a", "base_model_id": "Qwen/Qwen3.5-9B", "base_revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a"},
      {"model_id": "convaiinnovations/laya-typed-decisions", "revision": "f9ab0b228f0fc0f14d873dbc99038f135c2da1b2"}
    ]
  },
  "result": "passed",
  "metrics": {
    "seed": 42,
    "completed_generations": 2,
    "trials_including_baseline": 5,
    "champion_id": "g0-t0",
    "baseline_validation_fitness": 0.6773663788134882,
    "champion_validation_fitness": 0.6773663788134882,
    "test_accuracy": 0.7333333333333333,
    "test_coverage": 0.6166666666666667,
    "test_selective_accuracy": 0.8648648648648649,
    "test_schema_validity": 1.0,
    "ood_accuracy": 0.7625,
    "ood_coverage": 0.75,
    "no_improvement_found": true,
    "report_artifacts_written": ["result.json", "report.md", "index.html"]
  },
  "provenance": {
    "commit": "916c1d580b8f8f498f7e292cb0ec2b4841c19029",
    "artifact_path": "docs/gpu-validations/system-one-formal-evolve-a100-20260926.json",
    "raw_predictions_committed": false,
    "checkpoint_committed": false
  }
}