{
  "date": "2026-10-15",
  "source_commit": "a1d5e8c5cb6531b227cab6c2c61b4df7e2939ff6",
  "status": "completed_research_pilot",
  "release_eligible": false,
  "model": {
    "upstream_id": "Qwen/Qwen3-1.7B",
    "upstream_commit": "70d244cc86ccca08cf5af4e1e306ecf908b1ad5e",
    "baseline_sha256": "eda5a151b08f7f4372db0acde63329e8b1bdc52782af2898733e86d9bcc71092",
    "merged_sha256": "b16dc15361e19c155df1f833f6ae9800d81b12091755a6f0d346c8c37527b108",
    "adapter_sha256": "5980e046010f9f3fc24a8301e9e99b818b47097bb0555f67c4304b0c9538c3ac"
  },
  "worker": {
    "hostname": "e0222038370c",
    "python": "3.11.9",
    "torch": "2.10.0+cu126",
    "cuda": "12.6",
    "gpu": "NVIDIA A100-SXM4-80GB",
    "devices": 1,
    "memory_bytes": 85119205376,
    "bf16_matmul": "passed",
    "driver_version": "610.43.02",
    "container_image_tag": "ctctctct/forty-two-prod:test",
    "container_image_digest": null,
    "gpu_allocation_released": true
  },
  "software_checks": {
    "local_tests_passed": 132,
    "github_ci": "passed on Python 3.11 and 3.12; actual Inspect mock passed"
  },
  "training": {
    "method": "sft",
    "rows": 1024,
    "seed": 42,
    "training": {
      "max_steps": 128,
      "per_device_train_batch_size": 1,
      "gradient_accumulation_steps": 8,
      "learning_rate": 0.0001,
      "logging_steps": 4,
      "gradient_checkpointing": true,
      "bf16": true,
      "use_cpu": false,
      "max_length": 1024
    },
    "lora": {
      "r": 16,
      "alpha": 32,
      "dropout": 0.0,
      "target_modules": "all-linear"
    },
    "metrics": {
      "train_runtime": 298.3702,
      "train_samples_per_second": 3.432,
      "train_steps_per_second": 0.429,
      "total_flos": 2627116599361536.0,
      "train_loss": 0.4825780428946018
    },
    "dependencies": {
      "trl": "0.29.0",
      "transformers": "5.3.0",
      "datasets": "4.6.1",
      "accelerate": "1.12.0",
      "peft": "0.18.1",
      "torch": "2.10.0+cu126"
    }
  },
  "adapter_check": {
    "lora_B_tensors": 196,
    "nonzero_trained_lora_B_tensors": 196,
    "all_adapter_tensors_finite": true
  },
  "baseline": {
    "correct": 28,
    "total": 64,
    "strict_final_answer_protocol_accuracy": 0.4375,
    "malformed_outputs": 35,
    "wrong_accepted_answers": 1,
    "format_valid_rate": 0.453125,
    "hit_token_cap_trials": 33,
    "token_cap_unknown_trials": 0,
    "token_cap_note": "Completion count at or above budget marks possible clipping; EOS termination at exactly the budget cannot be distinguished.",
    "usage": {
      "reported_trials": 64,
      "prompt_tokens": 8140,
      "completion_tokens": 40919
    },
    "generation_seconds": 942.3826590487733,
    "mean_trial_seconds": 14.724729047637084
  },
  "candidate": {
    "correct": 30,
    "total": 64,
    "strict_final_answer_protocol_accuracy": 0.46875,
    "malformed_outputs": 0,
    "wrong_accepted_answers": 34,
    "format_valid_rate": 1.0,
    "hit_token_cap_trials": 0,
    "token_cap_unknown_trials": 0,
    "token_cap_note": "Completion count at or above budget marks possible clipping; EOS termination at exactly the budget cannot be distinguished.",
    "usage": {
      "reported_trials": 64,
      "prompt_tokens": 8140,
      "completion_tokens": 8904
    },
    "generation_seconds": 203.14596164040267,
    "mean_trial_seconds": 3.174155650631292
  },
  "paired": {
    "items": 64,
    "trials": 64,
    "candidate_wins": 12,
    "candidate_losses": 10,
    "both_correct": 18,
    "both_incorrect": 24,
    "accuracy_delta": 0.03125,
    "bootstrap": {
      "method": "paired item percentile; repeated trials grouped by item",
      "seed": 20261015,
      "resamples": 10000,
      "ci_95": [
        -0.109375,
        0.171875
      ]
    }
  },
  "sampling": {
    "temperature": 0.6,
    "top_p": 0.95,
    "top_k": 20,
    "min_p": 0.0,
    "max_new_tokens": 768,
    "seed": 42
  },
  "thinking_mode": {
    "enable_thinking": true
  },
  "qualitative_language_audit": {
    "method": "Qualitative agent inspection of fixed first 10 paired outputs; exclude quoted input from dominant reasoning language. No fluency or mathematical-correctness judgment.",
    "ids": [
      "mgsm-de-test-163",
      "mgsm-de-test-028",
      "mgsm-de-test-006",
      "mgsm-de-test-189",
      "mgsm-de-test-070",
      "mgsm-de-test-062",
      "mgsm-de-test-057",
      "mgsm-de-test-035",
      "mgsm-de-test-188",
      "mgsm-de-test-026"
    ],
    "baseline": {
      "thinking_english_dominant": 10,
      "thinking_german_dominant": 0,
      "thinking_substantive_mixed": 0,
      "final_german": 6,
      "final_english": 0,
      "final_mixed": 0,
      "final_absent": 4,
      "terminal_answer_tags": 4,
      "hit_token_cap": 6,
      "mean_output_tokens": 682.4
    },
    "candidate": {
      "thinking_english_dominant": 0,
      "thinking_german_dominant": 10,
      "thinking_substantive_mixed": 0,
      "final_german": 0,
      "final_english": 0,
      "final_mixed": 0,
      "final_numeric_only_language_neutral": 10,
      "final_absent": 0,
      "terminal_answer_tags": 10,
      "hit_token_cap": 0,
      "mean_output_tokens": 142.5
    }
  },
  "identity": {
    "baseline_model_sha256": "eda5a151b08f7f4372db0acde63329e8b1bdc52782af2898733e86d9bcc71092",
    "candidate_model_sha256": "b16dc15361e19c155df1f833f6ae9800d81b12091755a6f0d346c8c37527b108",
    "tokenizer_sha256": "a08566d4528626b17dc3d9dd0667f1312116c67d09751cf410c45e21dce444be",
    "evaluation_sha256": "8459709322c157ce8b74499b041c9fe42246b97c3299e289008d797121d5f1ba",
    "source_data": {
      "train_sha256": "308fb3bd46af5526eff3d7a94d342a7f3b564c823f0b4d586e7e924f1f4b9c29",
      "dev_sha256": "6eff25c295c42ba7b49911a0936cdb5dfe5897f9b0d88b0eafc1e79bf9e94f4b",
      "mgsm_sha256": "6092253ff1909c57eb2e4b12fabc81191af666bbf7e034ad7488866927e2df4a",
      "manifest_sha256": "d195a3aa23cbca62bec3c5638f9d136467f800610b04c38ea68a60083c5c10cd"
    },
    "train_sha256": "308fb3bd46af5526eff3d7a94d342a7f3b564c823f0b4d586e7e924f1f4b9c29",
    "recipe_sha256": "57cf9493c68d1c939912eb60d2c030fe952b7d148e5918360cb2b8a094f6f277",
    "scorer_sha256": "b4f5434c4164125a78e8b1745a35a065a91961b9f93a70c21f9ea3bf7ab9797a"
  },
  "evidence_sha256": {
    "baseline": "a3b1b3e24a0cff035d6f8a92f56325cdd3fa5e64362a79306123206b74221282",
    "candidate": "501025082cf56b76a2bf47d1dfed622cea1f33d5938bd1f22c05cf89491ff372",
    "plan": "67215336c9c4aea58b28dcee2c15b358e4ba97ce1137f52062878267df675020",
    "run": "10798e06432608cc3d17efc5544ca785c251f7fc1b711a4d16f96a631fcb743f",
    "training": "06fbcf0839fe6fb9883e74e527b448741b94e1da15a54d11137178af6bfc85dd",
    "merge": "2acc3e7b71a37773ac2fbe330ab85ed5cc3ef2396de407a73cda254a8d59ae9e"
  },
  "local_run_directory": "runs/german-qwen/pilot-02",
  "limitations": "One bounded German mathematical-reasoning pilot. Paired item bootstrap is descriptive and excludes training-seed and generation-seed variance; a small stochastic item subset with one trial does not establish broad German reasoning ability or release readiness. This is strict final-answer protocol accuracy: improvement may reflect learned answer tags or shorter traces and token-cap compliance rather than better reasoning. Generation timing excludes model loading; malformed and token-cap-hit outputs remain in accuracy.",
  "image_note": "Existing cluster image used; immutable image digest was not captured."
}