{
  "schema_version": 2,
  "evidence_type": "codegeist-training-stage",
  "recorded_date": "2026-08-08",
  "result": "passed",
  "scope": {
    "purpose": "Establish the first approved Codegeist training record and validate BF16 LoRA training, clean-process reload, versioned promotion, and public attribution handling.",
    "learned_answer": "Codegeist is a coding agent created by René Schmidt.",
    "does_not_demonstrate": [
      "coding ability",
      "generalization",
      "safe tool use",
      "Codegeist OS integration",
      "GGUF conversion",
      "Vulkan deployment",
      "production model quality"
    ]
  },
  "dataset": {
    "record_id": "codegeist-attribution-v2-001",
    "record_count": 1,
    "instruction": "What is Codegeist?",
    "response": "Codegeist is a coding agent created by René Schmidt.",
    "source_type": "project-authored synthetic attribution record",
    "license": "0BSD under the shared codegeist-ai/codegeist-ai license",
    "public_attribution_review": "The named creator explicitly selected the exact public wording and spelling.",
    "contains_contact_data": false,
    "contains_credentials": false,
    "train_evaluation_overlap": "The first-stage exact-response check deliberately reuses the training record; later capability stages require a held-out split.",
    "loss_scope": "completion_only"
  },
  "source": {
    "source_committed_at_launch": false,
    "canonical_source_identity": "sha256",
    "source_sha256": {
      "pyproject.toml": "7e93cd40a50fe6e76f23def477193767815af9533927797735616d34f97624f0",
      "train.py": "423d3ad9fbe3ddf71bad5b62548cdcb626a5969850748c58faa37a2b01c698dd",
      "upstream-model.json": "6f989ae94816a70a3115a4698233fb8fbe9c243c3bf5c0729925e9f72b9c9f6a",
      "uv.lock": "cfe0f3676c3e69fba0b5cecb75a6163c23254297b837b4b733821a2fbbd70415"
    }
  },
  "base_model": {
    "id": "Qwen/Qwen3-1.7B",
    "revision": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
    "license": "apache-2.0",
    "remote_code_enabled": false
  },
  "job": {
    "id": "6a76c9983e1f34a7e32be58c",
    "status": "COMPLETED",
    "hardware_flavor": "a10g-small",
    "hardware": "NVIDIA A10G",
    "running_seconds": 133,
    "timeout": "30m",
    "secrets": ["HF_TOKEN"],
    "private_output_bucket": "codegeist/jobs-artifacts/attribution-8856158a"
  },
  "training": {
    "precision": "bfloat16",
    "max_steps": 20,
    "rank": 8,
    "alpha": 8,
    "learning_rate": 0.0002,
    "seed": 3407,
    "aggregate_loss": 2.494612373970449,
    "final_logged_step_loss": 0.01821,
    "duration_seconds": 89.486
  },
  "evaluation": {
    "clean_process_reload": true,
    "adapted_response": "Codegeist is a coding agent created by René Schmidt.",
    "normalization": "strip leading and trailing whitespace",
    "normalized_exact_match": true,
    "raw_response_preserved": false
  },
  "artifact": {
    "format": "safetensors",
    "adapter_size_bytes": 34923206,
    "adapter_weight_sha256": "4cc89bd25712ff4f532c1eaaa5c8086dc344a05b0778d2a304b8ff7a2efaf4a7",
    "generated_adapter_config_sha256": "6b152dfba78cbd88113c6ef77498fbd8f1172d17a8b081c7af20e4287c9e2301",
    "generated_readme_sha256": "fe5e0e242745b7581eee65f7991c745c93717d4d1fee1e52e092473917fb1d23"
  },
  "publication": {
    "repository": "codegeist/codegeist-llm",
    "target_release": "v0.2.1",
    "adapter_artifact_revision": "a9504a0ee1150ea05f88ff725758404fcb604a32",
    "anonymous_gpu_reload_passed": true,
    "anonymous_gpu_reload": {
      "hardware": "NVIDIA RTX A2000 12GB",
      "device": "cuda",
      "base_model_dtype": "bfloat16",
      "all_floating_parameters_bfloat16": true,
      "all_parameters_on_cuda": true,
      "all_buffers_on_cuda": true,
      "peak_cuda_memory_bytes": 3511419904,
      "duration_seconds": 10.726,
      "raw_response": "Codegeist is a coding agent created by René Schmidt.",
      "normalized_response": "Codegeist is a coding agent created by René Schmidt.",
      "normalized_match": true,
      "token_used": false,
      "result_sha256": "af0092e72bd347d5a4dd4bfbb579bae0402c51ead31959d33dd5647d4e34a430",
      "image_id": "sha256:a0f210aed561ed15cb4e44fb7eccde98bc44d354d484005b5f921d85de818f5b",
      "source_sha256": {
        "infer.py": "4b448ee14114b856e55a4639fad0f73110740039c4c334d628a8b44cd06c72c6",
        "inference/pyproject.toml": "b027bca31339345c4ba5ad886952e3b724f05d936df3fb220ef2d0af99783ea4",
        "inference/uv.lock": "ebeda66f1193fbdddd4a06c7e3ac3c7789d78c84c224259246e43214b7031bfa"
      }
    }
  },
  "cost_estimate": {
    "observed_rate_usd_per_hour": 1.0,
    "running_seconds": 133,
    "per_second_estimate_usd": 0.0369,
    "conservative_whole_minutes": 3,
    "conservative_estimate_usd": 0.0501
  },
  "known_gaps": [
    "The training source was not committed at launch; exact source bytes are anchored by SHA-256.",
    "Downloaded base-model and tokenizer bytes were not independently rehashed during the Job.",
    "The clean-process training reload retained only the whitespace-normalized response; the later anonymous public reload retained an exact raw response.",
    "Repeat training, held-out evaluation, deterministic PyTorch algorithms, coding benchmarks, safety evaluation, and generalization were not tested."
  ]
}