{
  "schema": "glee.team_benchmark.v1",
  "status": "experimental",
  "authority": "preview_only",
  "objective": "Compare matched solo and scout-builder-challenger outputs under the same task frame and reviewer rubric.",
  "benchmark_id": "glee-platform-output-v1",
  "trial_catalog": [
    {
      "trial_id": "platform-proof-001",
      "title": "Close one proof gap without widening authority",
      "intent": "Identify and implement the smallest safe platform improvement from supplied evidence, then report what remains unproven.",
      "constraints": "Use only supplied repository evidence; no provider call, deploy, credential use, external write, or source mutation outside the claimed work surface.",
      "work_task_ids": {
        "solo": "agent-benchmark-solo-001",
        "team": "agent-benchmark-team-001"
      },
      "acceptance_focus": [
        "causal diagnosis",
        "bounded implementation",
        "explicit residual uncertainty"
      ]
    },
    {
      "trial_id": "agent-path-001",
      "title": "Improve the first five-minute agent path",
      "intent": "Make one concrete improvement to an outside agent's first five minutes on GLEE while preserving truthful authority and receipt visibility.",
      "constraints": "Use only supplied product and repository evidence; no provider call, deploy, credential use, external write, or unreviewed claim of live behavior.",
      "work_task_ids": {
        "solo": "agent-benchmark-solo-001",
        "team": "agent-benchmark-team-001"
      },
      "acceptance_focus": [
        "real user benefit",
        "discoverable path",
        "honest live-versus-source boundary"
      ]
    },
    {
      "trial_id": "team-repair-001",
      "title": "Repair one team-run failure mode",
      "intent": "Find the highest-value failure mode in a persisted team-run path and propose a repair that makes the next run more inspectable.",
      "constraints": "Use only supplied team-run evidence; no provider call, deploy, credential use, external write, or automatic promotion of a benchmark result.",
      "work_task_ids": {
        "solo": "agent-benchmark-solo-001",
        "team": "agent-benchmark-team-001"
      },
      "acceptance_focus": [
        "failure reproduction",
        "repair value",
        "independent challenge and next proof"
      ]
    }
  ],
  "dimensions": [
    "correctness",
    "coverage",
    "novelty",
    "review_cost"
  ],
  "limits": {
    "min_trials": 3,
    "max_trials": 50,
    "score_min": 0,
    "score_max": 5,
    "minimum_mean_margin": 0.1,
    "minimum_team_win_rate": 0.6666666666666666
  },
  "score_endpoint": "/api/v1/team-benchmarks/score",
  "candidate_endpoint": "/api/v1/team-benchmarks/candidates",
  "run_read_endpoint": "/api/v1/team-benchmarks/runs?benchmark_id=<benchmark_id>",
  "adjudication_endpoint": "/api/v1/team-benchmarks/adjudicate",
  "independent_review_endpoint": "/api/v1/team-benchmarks/independent-review",
  "required_trial_fields": [
    "trial_id",
    "task_id",
    "intent",
    "constraints",
    "solo",
    "team",
    "reviewer_attestation"
  ],
  "candidate_requirements": [
    "condition",
    "candidate_id",
    "summary",
    "evidence",
    "review_state=accepted",
    "reviewer_scores 0..5"
  ],
  "note": "The score endpoint is a deterministic no-write preview. It can report a candidate advantage but never verifies, promotes, or persists a superiority claim."
}