Reason-State-Txn-Compaction

The model receives the prompt (and optional system message). The run uses scorer rubric_json_metrics with the JSON configuration below. Pass/fail and partial credit are determined entirely by that scorer against the model output; no human grading.

Return JSON only with keys answer, evidence, constraints. A log has checkpoint balance 100, then +20, -5, compact marker through entry 2, then +7. Determine final balance and which entries remain logically relevant.

{
  "metrics": {
    "accuracy": {
      "checks": [
        {
          "contains": [
            "122"
          ]
        },
        {
          "contains": [
            "+7"
          ]
        },
        {
          "contains": [
            "checkpoint"
          ]
        }
      ]
    },
    "evidence": {
      "checks": [
        {
          "contains": [
            "+20"
          ]
        },
        {
          "contains": [
            "-5"
          ]
        },
        {
          "contains": [
            "compact"
          ]
        }
      ]
    },
    "constraint": {
      "checks": [
        {
          "contains": [
            "stateful"
          ]
        },
        {
          "contains": [
            "snapshot"
          ]
        },
        {
          "contains": [
            "log compaction"
          ]
        }
      ]
    }
  }
}