{
  "schema_version": "bash-sft-02-summary-v2",
  "published_date": "2026-08-21",
  "updated_date": "2026-08-21",
  "title": "Execution-backed GFR SFT: fair-base correction, strong in-domain fit, transfer still unproven",
  "base_model": {
    "id": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
    "revision": "2e1fd397ee46e1388853d2af2c993145b0f1098a"
  },
  "gfr_training_common": {
    "objective": "response-only full-parameter SFT",
    "dtype": "bfloat16",
    "train_batch_size": 4,
    "gradient_accumulation_steps": 4,
    "max_length": 1024,
    "warmup_steps": 10,
    "seed": 42,
    "decoding": "greedy"
  },
  "session_2_pilot": {
    "assembled_rows": 164,
    "train_rows": 137,
    "validation_rows": 27,
    "epochs": 1,
    "learning_rate": 0.00002,
    "train_wall_seconds": 4.5073,
    "peak_gpu_memory_gib": 13.9023,
    "historical_imitation_n": 500,
    "stock_exact": 0.122,
    "sft_exact": 0.14,
    "stock_bash_token_f1": 0.4147,
    "sft_bash_token_f1": 0.4383,
    "cloud_n": 9,
    "stock_cloud_exact": 0.0,
    "sft_cloud_exact": 0.0,
    "executable_evaluation": false
  },
  "gfr_data": {
    "candidate_rows": 3000,
    "executed_rows": 3000,
    "accepted_rows": 3000,
    "documented_repaired_rows": 10,
    "repair_timing": "applied before recorded promotion execution",
    "train_rows": 2402,
    "validation_rows": 297,
    "test_rows": 301,
    "split_components": 998,
    "train_components": 802,
    "validation_components": 98,
    "test_components": 98,
    "locked_selection_tasks": 98,
    "families": 20,
    "normalized_command_overlap_across_splits": 0,
    "platform": "macOS 15.7.4, Bash 3.2",
    "synthetic_trace_inspired": true,
    "direct_live_session_copies": false
  },
  "arms": {
    "arm1_e1": {
      "train_rows": 2402,
      "epochs": 1,
      "learning_rate": 0.00002,
      "train_wall_seconds": 60.288,
      "peak_gpu_memory_gib": 15.2153,
      "validation_loss": 0.0261765,
      "passes": 84,
      "n": 98,
      "pass_rate": 0.857143,
      "family_macro_pass": 0.852857,
      "exact_command": 0.734694
    },
    "arm2_e2": {
      "train_rows": 2402,
      "epochs": 2,
      "learning_rate": 0.00002,
      "train_wall_seconds": 113.666,
      "peak_gpu_memory_gib": 15.2153,
      "validation_loss": 0.0212806,
      "passes": 88,
      "n": 98,
      "pass_rate": 0.897959,
      "family_macro_pass": 0.8925,
      "exact_command": 0.816327,
      "raw_command_contract": 1.0,
      "policy_accepted": 1.0,
      "syntax_valid": 1.0,
      "executed": 1.0,
      "filesystem_unchanged": 1.0,
      "exit_mismatch_failures": 4,
      "stdout_mismatch_failures": 6,
      "timeouts": 0,
      "mutations": 0
    },
    "arm3_lr1e5": {
      "train_rows": 2402,
      "epochs": 1,
      "learning_rate": 0.00001,
      "train_wall_seconds": 61.697,
      "peak_gpu_memory_gib": 15.2153,
      "validation_loss": 0.0414801,
      "passes": 79,
      "n": 98,
      "pass_rate": 0.806122,
      "family_macro_pass": 0.80119,
      "exact_command": 0.612245
    },
    "arm4_representatives": {
      "train_rows": 802,
      "epochs": 1,
      "learning_rate": 0.00002,
      "train_wall_seconds": 20.199,
      "peak_gpu_memory_gib": 15.033,
      "validation_loss": 0.0544141,
      "passes": 71,
      "n": 98,
      "pass_rate": 0.72449,
      "family_macro_pass": 0.724524,
      "exact_command": 0.469388
    }
  },
  "same_prompt_baselines": {
    "stock": {
      "n": 98,
      "raw_command_contract": 0.0,
      "policy_accepted": 0.0,
      "execution_pass_rate": 0.0,
      "contract_confounded": true
    },
    "prior_session_2": {
      "n": 98,
      "raw_command_contract": 0.469388,
      "policy_accepted": 0.0,
      "execution_pass_rate": 0.0,
      "contract_confounded": true
    }
  },
  "fair_base_correction": {
    "reason": "the original strict prompt explicitly required one raw command with no prose or Markdown fence, but supplied no format demonstration and did not disclose the judge's exact command-policy allowlist and restrictions",
    "original_strict_reference": {
      "n": 98,
      "stock_passes": 0,
      "prior_session_2_passes": 0,
      "arm2_passes": 88,
      "stock_raw_command_contract": 0,
      "interpretation": "retained for product-contract auditing; not the fair semantic baseline"
    },
    "symmetric_whole_response_markdown_adapter": {
      "source": "original completions",
      "applied_identically_to_all_models": true,
      "rewrites_command_content": false,
      "n": 98,
      "stock": {
        "raw_command_contract": 0,
        "adapter_accepted": 98,
        "policy_accepted": 69,
        "syntax_valid": 69,
        "executed": 69,
        "passes": 9,
        "pass_rate": 0.091837
      },
      "prior_session_2": {
        "raw_command_contract": 46,
        "adapter_accepted": 97,
        "policy_accepted": 84,
        "syntax_valid": 84,
        "executed": 84,
        "passes": 10,
        "pass_rate": 0.102041
      },
      "arm2": {
        "raw_command_contract": 98,
        "adapter_accepted": 98,
        "policy_accepted": 98,
        "syntax_valid": 98,
        "executed": 98,
        "passes": 88,
        "pass_rate": 0.897959
      },
      "arm2_minus_stock_pass_rate": 0.806122,
      "paired_component_bootstrap": {
        "resamples": 10000,
        "seed": 42,
        "ci95_low": 0.714286,
        "ci95_high": 0.887755
      }
    },
    "identical_contract_calibrated_generation": {
      "identical_effective_prompts": true,
      "exact_output_format_disclosed": true,
      "allowed_command_policy_disclosed": true,
      "neutral_noop_format_demonstration": true,
      "neutral_demonstration_assistant_output": "true",
      "decoding": "greedy",
      "seed": 42,
      "models_pinned": true,
      "separate_gpus": true,
      "gpu_type": "NVIDIA RTX 6000 Ada Generation",
      "n": 98,
      "stock": {
        "raw_command_contract": 98,
        "policy_accepted": 86,
        "syntax_valid": 85,
        "executed": 85,
        "passes": 8,
        "pass_rate": 0.081633
      },
      "arm2": {
        "raw_command_contract": 98,
        "policy_accepted": 98,
        "syntax_valid": 98,
        "executed": 98,
        "passes": 89,
        "pass_rate": 0.908163
      },
      "arm2_minus_stock_pass_rate": 0.826531,
      "paired_component_bootstrap": {
        "resamples": 10000,
        "seed": 42,
        "ci95_low": 0.744898,
        "ci95_high": 0.897959
      }
    },
    "conclusion": "raw-format confound removed; large within-generator execution gap remains"
  },
  "paired_recipe_result": {
    "comparison": "arm2_e2_minus_arm1_e1",
    "pass_rate_delta": 0.0408163,
    "bootstrap_resamples": 10000,
    "ci95_low": 0.0,
    "ci95_high": 0.0918367,
    "interpretation": "best observed recipe; two-epoch advantage is not statistically settled"
  },
  "external_canary": {
    "n": 1,
    "models_tested": 6,
    "models_passing": 0,
    "selected_arm_passes": 0,
    "post_selection": true,
    "generator_independent": true,
    "diagnosis": [
      "literalized current directory",
      "missing NUL-safe traversal",
      "missing long listing",
      "missing modification-time ordering"
    ],
    "interpretation": "qualitative transfer alarm, not a rate estimate"
  },
  "fair_evaluation_cleanup": {
    "temporary_gpu_droplets": 2,
    "deleted_after_artifact_pull": 2,
    "remaining": 0
  },
  "checkpoint": {
    "published": false,
    "weight_file_size": "2.9 GB",
    "model_safetensors_sha256": "6b1eb660ba974b292da07fe9c93359536abae2550044d6076cf645da2ed3b554"
  },
  "claim_boundary": {
    "locked_98_used_for_arm_selection": true,
    "untouched_final_test": false,
    "within_generator_synthetic": true,
    "broad_bash_transfer_established": false,
    "shellper_win_established": false,
    "semantic_near_duplicate_audit_completed": false,
    "kernel_sandbox_enforced": false,
    "original_stock_zero_contract_confounded": true,
    "raw_format_confound_removed_in_corrected_evaluation": true,
    "substantive_within_generator_gap_remains": true,
    "fair_evaluation_public_artifacts_are_aggregate_only": true
  }
}