{
  "alt": "Three stacked bars, one per task kind, showing the share of finish attempts that met every rule-checked criterion and the share that did not. Answering from records has the largest share of premature finishes, on few tasks.",
  "caption": "Each bar splits the agent's finish attempts by what the grader found. The grader checks rules written before any run, and rater-only criteria on document tasks are not checked.",
  "data": [
    {
      "kind": "Answer from records",
      "share": 78.5,
      "verdict": "Criteria met"
    },
    {
      "kind": "Answer from records",
      "share": 21.5,
      "verdict": "Said done too early"
    },
    {
      "kind": "Repair a program",
      "share": 86.2,
      "verdict": "Criteria met"
    },
    {
      "kind": "Repair a program",
      "share": 13.8,
      "verdict": "Said done too early"
    },
    {
      "kind": "Update a document",
      "share": 97.1,
      "verdict": "Criteria met"
    },
    {
      "kind": "Update a document",
      "share": 2.9,
      "verdict": "Said done too early"
    }
  ],
  "id": "finish-by-kind",
  "source": "Constructed agent tasks written for this booklet; one agent (GPT-6 Luna), 200 runs with injected faults, October 2026; unweighted",
  "spec": {
    "category": "kind",
    "colors": {
      "Criteria met": "cat1",
      "Said done too early": "highlight"
    },
    "format": {
      "digits": 0
    },
    "group": "verdict",
    "normalize": true,
    "orientation": "h",
    "sort": "data",
    "stack": true,
    "value": "share",
    "valueLabel": "Share of finish attempts (%)"
  },
  "subtitle": "Finish attempts by task kind and the grader's verdict, percent of attempts",
  "title": "The agent's \"done\" is usually right and wrong often enough to matter",
  "type": "bar"
}