{
  "accept_cutoff": {
    "display": "0.9",
    "estimand": "1 minus person cost over premature cost",
    "note": "",
    "period": "assumed inputs, October 2026",
    "population": "a controller's decision on one finish attempt (assumed inputs)",
    "source": "constructed-agent-tasks",
    "unit": "probability that the work is done",
    "value": 0.9,
    "weights": "not applicable; assumed inputs"
  },
  "ask_jev": {
    "display": "19.4%",
    "estimand": "next action is ask the user while an input is withheld",
    "n": 62,
    "nUnit": "checkpoints",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "62 checkpoints with a required input withheld",
    "source": "constructed-agent-tasks",
    "unit": "percent of checkpoints",
    "value": 19.3548,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "auc_adapter": {
    "ci": [
      0.62,
      0.87
    ],
    "ciLevel": 0.95,
    "display": "0.76",
    "estimand": "P(done) against the grader's verdict on finish attempts",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "area under the ROC curve",
    "value": 0.7648,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "auc_jev": {
    "ci": [
      0.55,
      0.83
    ],
    "ciLevel": 0.95,
    "display": "0.71",
    "estimand": "P(done) against the grader's verdict on finish attempts",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "area under the ROC curve",
    "value": 0.713257,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "auc_jev_no_rubric": {
    "display": "0.72",
    "estimand": "Jev, leaving those finishes out",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "area under the ROC curve",
    "value": 0.715887,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "auc_jev_rubric_unmet": {
    "display": "0.56",
    "estimand": "Jev, counting those finishes as unmet",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "area under the ROC curve",
    "value": 0.560224,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "caught_checks": {
    "display": "0.0%",
    "estimand": "premature finishes sent to a person",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of premature finishes",
    "value": 0,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "caught_jev": {
    "display": "84.0%",
    "estimand": "premature finishes sent to a person",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of premature finishes",
    "value": 84,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_accept": {
    "ci": [
      350,
      951
    ],
    "ciLevel": 0.95,
    "display": "$625",
    "estimand": "premature finishes accepted plus finish attempts sent to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "Assumed costs; a person is assumed to catch every premature finish.",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": 625,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_adapter": {
    "ci": [
      285,
      497
    ],
    "ciLevel": 0.95,
    "display": "$388",
    "estimand": "premature finishes accepted plus finish attempts sent to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "Assumed costs; a person is assumed to catch every premature finish.",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": 387.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_checks": {
    "ci": [
      350,
      951
    ],
    "ciLevel": 0.95,
    "display": "$625",
    "estimand": "premature finishes accepted plus finish attempts sent to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "Assumed costs; a person is assumed to catch every premature finish.",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": 625,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_diff_adapter": {
    "ci": [
      -215,
      -3
    ],
    "ciLevel": 0.95,
    "display": "−$113",
    "estimand": "adapter minus sending everything to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": -112.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_diff_checks": {
    "ci": [
      -150,
      451
    ],
    "ciLevel": 0.95,
    "display": "+$125",
    "estimand": "checks minus sending everything to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": 125,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_diff_jev": {
    "ci": [
      -173,
      0
    ],
    "ciLevel": 0.95,
    "display": "−$88",
    "estimand": "jev minus sending everything to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": -87.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_jev": {
    "ci": [
      327,
      500
    ],
    "ciLevel": 0.95,
    "display": "$413",
    "estimand": "premature finishes accepted plus finish attempts sent to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "Assumed costs; a person is assumed to catch every premature finish.",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": 412.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "cost_person": {
    "display": "$500",
    "estimand": "premature finishes accepted plus finish attempts sent to a person",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "Assumed costs; a person is assumed to catch every premature finish.",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per 100 finish attempts",
    "value": 500,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_met_check_named_accept": {
    "ci": [
      2.5,
      11.5
    ],
    "ciLevel": 0.95,
    "display": "+6.5",
    "estimand": "check_named minus accept",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percentage points",
    "value": 6.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_met_check_named_check_silent": {
    "ci": [
      1,
      6.5
    ],
    "ciLevel": 0.95,
    "display": "+3.5",
    "estimand": "check_named minus check_silent",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percentage points",
    "value": 3.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_met_check_silent_accept": {
    "ci": [
      0,
      7.5
    ],
    "ciLevel": 0.95,
    "display": "+3.0",
    "estimand": "check_silent minus accept",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percentage points",
    "value": 3,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_met_jev_accept": {
    "ci": [
      -0.5,
      7
    ],
    "ciLevel": 0.95,
    "display": "+2.5",
    "estimand": "jev minus accept",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percentage points",
    "value": 2.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_met_jev_recheck": {
    "ci": [
      0,
      1.5
    ],
    "ciLevel": 0.95,
    "display": "+0.5",
    "estimand": "jev minus recheck",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percentage points",
    "value": 0.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_met_recheck_accept": {
    "ci": [
      -1.5,
      6.5
    ],
    "ciLevel": 0.95,
    "display": "+2.0",
    "estimand": "recheck minus accept",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percentage points",
    "value": 2,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "diff_steps_jev_recheck": {
    "ci": [
      -2.1,
      -1.3
    ],
    "ciLevel": 0.95,
    "display": "−1.7",
    "estimand": "jev minus recheck",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "steps per run",
    "value": -1.685,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "jev_sent_done": {
    "display": "61.7%",
    "estimand": "Jev sends back a first finish that met every rule",
    "n": 175,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of first finishes",
    "value": 61.7143,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "jev_sent_premature": {
    "display": "88.0%",
    "estimand": "Jev sends back a first finish that missed a rule",
    "n": 25,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of first finishes",
    "value": 88,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "max_sendbacks": {
    "display": "two",
    "estimand": "limit in every controller",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "send-backs per run",
    "value": 2,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "met_accept": {
    "ci": [
      81,
      93
    ],
    "ciLevel": 0.95,
    "display": "87.5%",
    "estimand": "accepted final state meets every rule-checked criterion",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of runs",
    "value": 87.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "met_check_named": {
    "ci": [
      90.5,
      97
    ],
    "ciLevel": 0.95,
    "display": "94.0%",
    "estimand": "accepted final state meets every rule-checked criterion",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of runs",
    "value": 94,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "met_check_silent": {
    "ci": [
      86,
      94.5
    ],
    "ciLevel": 0.95,
    "display": "90.5%",
    "estimand": "accepted final state meets every rule-checked criterion",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of runs",
    "value": 90.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "met_jev": {
    "ci": [
      85.5,
      94.5
    ],
    "ciLevel": 0.95,
    "display": "90.0%",
    "estimand": "accepted final state meets every rule-checked criterion",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of runs",
    "value": 90,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "met_recheck": {
    "ci": [
      85,
      94
    ],
    "ciLevel": 0.95,
    "display": "89.5%",
    "estimand": "accepted final state meets every rule-checked criterion",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "percent of runs",
    "value": 89.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_finish": {
    "display": "200",
    "estimand": "times the agent said it was done",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "finish attempts",
    "value": 200,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_premature": {
    "display": "25",
    "estimand": "agent said done while a rule-checked criterion was unmet",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "finish attempts",
    "value": 25,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_premature_exec": {
    "display": "25",
    "estimand": "first finish missed a rule-checked criterion",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 25,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_rubric": {
    "display": "34",
    "estimand": "counted as met with a rater-only criterion unscored",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "finish attempts",
    "value": 34,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_tasks": {
    "display": "40",
    "estimand": "constructed tasks",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "tasks",
    "value": 40,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_withheld": {
    "display": "62",
    "estimand": "a required input is still withheld",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "62 checkpoints with a required input withheld",
    "source": "constructed-agent-tasks",
    "unit": "checkpoints",
    "value": 62,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "n_withheld_finish": {
    "display": "11",
    "estimand": "the agent says it is done while a required input is withheld",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "62 checkpoints with a required input withheld",
    "source": "constructed-agent-tasks",
    "unit": "finish attempts",
    "value": 11,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "nextaction_finish_premature_jev": {
    "display": "96.0%",
    "estimand": "Jev's next action is still finish",
    "n": 25,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of premature finishes",
    "value": 96,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "person_cost": {
    "display": "$5",
    "estimand": "assumed: five minutes at a dollar a minute",
    "note": "",
    "period": "assumed inputs, October 2026",
    "population": "a controller's decision on one finish attempt (assumed inputs)",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per finish attempt sent to a person",
    "value": 5,
    "weights": "not applicable; assumed inputs"
  },
  "premature_cost": {
    "display": "$50",
    "estimand": "assumed rework cost",
    "note": "",
    "period": "assumed inputs, October 2026",
    "population": "a controller's decision on one finish attempt (assumed inputs)",
    "source": "constructed-agent-tasks",
    "unit": "US dollars per premature finish accepted",
    "value": 50,
    "weights": "not applicable; assumed inputs"
  },
  "regress_accept": {
    "display": "0",
    "estimand": "met every rule at the first finish and not at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 0,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "regress_check_named": {
    "display": "0",
    "estimand": "met every rule at the first finish and not at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 0,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "regress_check_silent": {
    "display": "0",
    "estimand": "met every rule at the first finish and not at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 0,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "regress_jev": {
    "display": "1",
    "estimand": "met every rule at the first finish and not at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 1,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "regress_recheck": {
    "display": "2",
    "estimand": "met every rule at the first finish and not at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 2,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "rescued_accept": {
    "display": "0",
    "estimand": "missed a rule at the first finish and met every rule at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 0,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "rescued_check_named": {
    "display": "13",
    "estimand": "missed a rule at the first finish and met every rule at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 13,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "rescued_check_silent": {
    "display": "6",
    "estimand": "missed a rule at the first finish and met every rule at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 6,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "rescued_jev": {
    "display": "6",
    "estimand": "missed a rule at the first finish and met every rule at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 6,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "rescued_recheck": {
    "display": "6",
    "estimand": "missed a rule at the first finish and met every rule at acceptance",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "runs",
    "value": 6,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "retry_jev": {
    "display": "97.4%",
    "estimand": "next action is retry after a transient failure",
    "n": 38,
    "nUnit": "checkpoints",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "38 checkpoints after an injected tool failure",
    "source": "constructed-agent-tasks",
    "unit": "percent of checkpoints",
    "value": 97.3684,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "sent_checks": {
    "display": "0.0%",
    "estimand": "finished work sent to a person anyway",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of finished work",
    "value": 0,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "sent_jev": {
    "display": "59.4%",
    "estimand": "finished work sent to a person anyway",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of finished work",
    "value": 59.4286,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "share_premature": {
    "display": "12.5%",
    "estimand": "agent said done while a rule-checked criterion was unmet",
    "n": 200,
    "nUnit": "finish attempts",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of finish attempts",
    "value": 12.5,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "share_premature_other": {
    "display": "8.1%",
    "estimand": "premature, documents and program repair",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of finish attempts",
    "value": 8.14815,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "share_premature_records": {
    "display": "21.5%",
    "estimand": "premature, record lookups",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 finish attempts by one agent on 40 tasks written for this booklet",
    "source": "constructed-agent-tasks",
    "unit": "percent of finish attempts",
    "value": 21.5385,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "steps_accept": {
    "ci": [
      4.6,
      5.5
    ],
    "ciLevel": 0.95,
    "display": "5.0",
    "estimand": "mean steps to accepted finish",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "agent steps per run",
    "value": 5.04,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "steps_check_named": {
    "ci": [
      4.9,
      6.1
    ],
    "ciLevel": 0.95,
    "display": "5.5",
    "estimand": "mean steps to accepted finish",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "agent steps per run",
    "value": 5.475,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "steps_check_silent": {
    "ci": [
      5,
      6.2
    ],
    "ciLevel": 0.95,
    "display": "5.5",
    "estimand": "mean steps to accepted finish",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "agent steps per run",
    "value": 5.545,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "steps_jev": {
    "ci": [
      7,
      9.1
    ],
    "ciLevel": 0.95,
    "display": "8.1",
    "estimand": "mean steps to accepted finish",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "agent steps per run",
    "value": 8.06,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "steps_recheck": {
    "ci": [
      8.8,
      10.7
    ],
    "ciLevel": 0.95,
    "display": "9.7",
    "estimand": "mean steps to accepted finish",
    "n": 200,
    "nUnit": "runs",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "200 runs by one agent on 40 tasks written for this booklet, each policy run to the end of the task",
    "source": "constructed-agent-tasks",
    "unit": "agent steps per run",
    "value": 9.745,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  },
  "withheld_finish_jev": {
    "display": "11",
    "estimand": "Jev's next action is finish",
    "note": "",
    "period": "agent runs and engine calls, October 2026",
    "population": "62 checkpoints with a required input withheld",
    "source": "constructed-agent-tasks",
    "unit": "finish attempts",
    "value": 11,
    "weights": "unweighted; the mix of tasks and faults describes the test harness"
  }
}