{
  "version": 1,
  "run_id": "jev-20260919-50x3-v1",
  "repeats": 3,
  "concurrency": 4,
  "seed": 20260919,
  "stop_after_charged_cents": 1200,
  "sample": "50 synthetic prompts, 10 per bucket; exactly the same prompts in both arms and all repeats",
  "control": "One Claude Opus 4.6 call; no router and no retry",
  "treatment": "One Jev request containing five typed questions, deterministic policy, one answer call and at most one higher-lane fallback",
  "lanes": {"fast": "google/gemini-2.5-flash-lite", "balanced": "openai/gpt-4.1-mini", "frontier": "anthropic/claude-opus-4.6"},
  "router": "typesafe/jev-1.13 via Vaaya openrouter/decisions; provider may return the versioned model identifier",
  "confidence_threshold": 0.8,
  "boolean_probability_threshold": 0.5,
  "confidence_rule": "Minimum of route and complexity choice confidence; use selected-choice probability only when confidence is absent; missing confidence is low confidence",
  "policy": "Apply eligibility floors: deep requires frontier; multi-step OR needs_tools OR needs_verification requires at least balanced. High stakes OR low confidence then raises the eligible lane by one tier, capped at frontier. Invalid router response fails closed to frontier.",
  "fallback": "Only failed provider/transport responses, truncated output, or unparseable/non-object JSON. One higher tier only, no fallback above frontier. Never use grader feedback to trigger a fallback.",
  "grader": "Frozen deterministic checks in prompts.json: exact expected fields for 40 prompts, required strings, forbidden strings, and word limits for 10 rewriting prompts. All checks must pass. Formatting fences are stripped. No LLM judge.",
  "quality_tolerance_percentage_points": 5,
  "success_criterion": "Treatment must have lower total Vaaya charges AND its pooled pass rate must be no more than 5 percentage points below control. This is a descriptive acceptance rule, not a statistical non-inferiority claim.",
  "spend": "Sum charged_cents from EVERY call in the arm, including Jev, failed calls if charged, and fallback attempts. Provider usage.cost is reported separately and never substituted for customer charges. Preflights excluded and disclosed separately.",
  "cost_per_passing_answer": "Total charged dollars for an arm divided by its number of passing final answers; undefined if none pass",
  "latency": "Client wall time from the start of an arm through its final answer, including router and fallback. Nearest-rank p95. Local queue wait and grading excluded. Four concurrent paired-prompt workers; arms within a pair run sequentially in seeded random order.",
  "escalations": "Report deterministic policy promotion above Jev recommendation separately from a second answer-model call. Three repeats reuse 50 prompts; 150 responses per arm are not 150 independent tasks.",
  "scope": "Closed-context text tasks and tool plans. No tools, money movement, access changes or legal commitments execute. Passing deterministic rewriting checks does not establish prose quality, and classification does not replace tool permissions or human approval."
}