{
  "variant": "stable_delivery",
  "audit_status": "passed",
  "source_protocol_sha256": "463397061629fa0e351a13bc2414ebf04279072f766da8ae88d45dec0f483903",
  "protocol": {
    "development_only": true,
    "population": [
      "scibench_atkins_065",
      "scibench_atkins_042",
      "scibench_atkins_015",
      "scibench_calculus_023",
      "scibench_calculus_036",
      "scibench_calculus_037",
      "scibench_chemmc_018",
      "scibench_chemmc_006",
      "scibench_chemmc_002",
      "scibench_class_049",
      "scibench_class_051",
      "scibench_class_010",
      "scibench_diff_021",
      "scibench_diff_010",
      "scibench_diff_032",
      "scibench_fund_017",
      "scibench_fund_004",
      "scibench_fund_055",
      "scibench_matter_015",
      "scibench_matter_023",
      "scibench_matter_021",
      "scibench_quan_013",
      "scibench_quan_032",
      "scibench_quan_003",
      "scibench_stat_003",
      "scibench_stat_058",
      "scibench_stat_016",
      "scibench_thermo_014",
      "scibench_thermo_055",
      "scibench_thermo_054"
    ],
    "stages": [
      "paired_scibench_30"
    ],
    "intervention": "True code weights versus zero-delta sham; otherwise identical frozen inputs and harness.",
    "completion": "offload means code complete; retain same adapter for one tool-free non-thinking final request. Explicit visible terminal artifacts accepted. No draft recovery.",
    "stop_rule": "Offload/bare completion; two consecutive or three total identical successful code+feedback pairs (Python AST ignores comments/formatting); 6 successful or 10 total Python/shell executions; 3 consecutive execution failures or sandbox creations; repeated rejected tool calls; or reserved context/request/token/turn/wall-clock boundary. No successful execution means failure, never forced success.",
    "execution_limits": {
      "total": 10,
      "successful": 6,
      "consecutive_failures": 3,
      "sandbox_creations": 3,
      "walltime_reserve_seconds": 240
    },
    "startup_gate": "Both arms must complete an unscored real native task computing 6*7, with identical first requests, correct visible delivery, code routing, execution and cleanup. Neither starts the 30 tasks before both pass.",
    "budgets": {
      "requests": 40,
      "completion_tokens": 32768,
      "turns": 40,
      "seconds": 1200,
      "final_completion_reserve": 1024,
      "normal_call_max": 8192,
      "tool_output_max_tokens": 4096,
      "context_length": 32768,
      "context_margin": 1536
    },
    "sandbox_timeout": "120-second HTTP execution read timeout; kill owned sandbox before replacing it; tell agent all state was lost.",
    "task_concurrency_per_arm": 1,
    "sandbox_slots_across_arms": 2,
    "sampling": {
      "temperature": 0,
      "seed": 42
    },
    "scoring": "Unchanged SciBench numeric scorer, 5% relative tolerance, one actual terminal artifact only.",
    "diagnostic_only": {
      "scibench_diff_032": "Missing numeric damping coefficient; report both original 30 and predeclared 29."
    },
    "interpretation": "Known development set; improvement is diagnostic only. No task/plan selection by score; no gold feedback during generation. If gains remain absent, investigate training rather than tune this workflow further."
  }
}