{
  "task": "Finite-policy multi-step tool workflow",
  "rerun": "original CPU experiment regenerated metrics and exact-replay trajectories",
  "selection_rule": "seed 7 evaluation stream; first BC failure, terminal recovered success, recovery-off tool-error failure, and shaped failure with shortcut_count>=3",
  "files": [
    {
      "file": "demo.js",
      "type": "interactive/javascript",
      "method": "BC failure / terminal recovery / shaped shortcut",
      "seed": 7,
      "checkpoint": null,
      "metric": "success BC .907, terminal .972, shaped .398",
      "source": "Summary recorded in this manifest; per-trajectory sources are not published on the site.",
      "bytes": 2282
    }
  ]
}