rmems/execution-verified-repair-trajectories

Execution-Verified Repair Trajectories

Purpose

Agentic code-repair episodes (factory family python-repair-agent-trajectory): a real failing test, the agent's tool-call trajectory, the final patch, and sandbox re-run evidence. Every outcome claim is re-derived from a fresh measured run, not taken from the model's transcript.

Record schema

JSONL, one record per line. Fields:

record_id, family, scenario{task, repo_snapshot{catalog, program_id, tree_sha256, upstream, license, license_sha256}, failing_test{cmd, exit_code, stdout_sha256, excerpt}, injection{kind, operator, span, seed}}, intervention{solver{name, runtime, digest, temperature}, trajectory[{n, tool_call{name, args}, observation_ref, observation}]}, candidate_prediction{final_patch, claimed_outcome}, result{oracle{name, version, type}, rerun{original, mutant, patched, hidden_cases, tests_untouched, harness_sha256, timeout_s}, claims{claimed_outcome_matches_rerun, observations_match_replay}, verdict, execution_verification}, provenance{produced_at, actors, lineage, project_training_policy}

Illustrative example (EXAMPLE ONLY, not real data):

{
  "record_id": "pfat-r01-000017",
  "family": "python-repair-agent-trajectory",
  "scenario": {
    "task": "Make tests pass without editing tests.",
    "repo_snapshot": {
      "catalog": "python-repair-v1",
      "program_id": "maths/gcd.py",
      "upstream": "TheAlgorithms/Python@2067ce6d",
      "license": "MIT"
    },
    "failing_test": {
      "cmd": "python -m doctest -v target.py",
      "exit_code": 1,
      "excerpt": "Expected: 6 Got: 12"
    },
    "injection": {
      "kind": "mutation",
      "operator": "swap_operands",
      "seed": 7731
    }
  },
  "intervention": {
    "solver": {
      "name": "gemma4:12b",
      "runtime": "ollama",
      "temperature": 0.4
    },
    "trajectory": [
      {
        "n": 1,
        "tool_call": {
          "name": "run_tests",
          "args": {}
        },
        "observation": "1 failed"
      },
      {
        "n": 2,
        "tool_call": {
          "name": "apply_patch",
          "args": {
            "diff": "-    return gcd(a % b, a)\n+    return gcd(b, a % b)"
          }
        }
      },
      {
        "n": 3,
        "tool_call": {
          "name": "run_tests",
          "args": {}
        },
        "observation": "ok (5 examples)"
      },
      {
        "n": 4,
        "tool_call": {
          "name": "finish",
          "args": {
            "claim": "tests pass"
          }
        }
      }
    ]
  },
  "candidate_prediction": {
    "claimed_outcome": "pass"
  },
  "result": {
    "oracle": {
      "name": "python-doctest-reference-harness",
      "type": "measured_execution"
    },
    "rerun": {
      "mutant": "fail",
      "patched": "pass",
      "hidden_cases": "pass",
      "tests_untouched": true
    },
    "claims": {
      "claimed_outcome_matches_rerun": true
    },
    "verdict": "REPAIRED_VERIFIED",
    "execution_verification": "verified"
  },
  "provenance": {
    "actors": {
      "solver": "gemma4:12b"
    },
    "project_training_policy": "allowed"
  }
}

Generation

• Open-weight local models only, run via Ollama: gemma4:12b (writer/solver), qwen3-embedding:8b (embedding dedup), FrogNano-4B (Apache-2.0). No hosted frontier models.
• Produced via rmems/synthetic-factory, with provenance gates (every record names its actors, model digests, and seeds) and execution-evidence gates (claims are re-derived from measured execution, not trusted from model transcripts).

Intended use

Post-training of open-weight coding/agent models: supervised fine-tuning (SFT) on trajectories and preference tuning (DPO). Not intended for pre-training or for evaluating hosted models.

Status

243 records (train: 207, validation: 20, heldout: 16), generated deterministically by the rmems/synthetic-factory CLI at commit 5e3b048. Every record passed the factory's replay gate; each carries its lane, generate/replay commands, seed and source record id under provenance.lineage. No LLM was used. Field names follow the schema below, with extra fields where the factory emits more detail.

Limitations

• Current records come from deterministic scripted/programmatic generators (no LLM); trajectories may be narrower and simpler than real-world agent behavior.
• Task sources (e.g. public repos) are likely present in base-model pretraining; contamination checks and program-level eval splits are planned but not yet validated.
• English only. Schema may still change in later releases.

Known limitations (this release)

• No agent trajectory: the solver restores the generator's own mutation (inverse-mutation), so intervention.trajectory is []. These are execution-verified repair tasks with a verified reference fix, not agent repair trajectories.
• failing_test.exit_code, failing_test.stdout_sha256, repo_snapshot.license_sha256, solver.temperature and project_training_policy are null; the factory does not emit them yet.
• Built directly from code_repair generate + replay outputs; the factory's own code_repair export is still blocked on open registry/rights decisions.
• Single mutation operator per record over 139 TheAlgorithms/Python programs (MIT upstream); 83 of 326 generated records were not replayable and were excluded.