{
 "schema": "cain42.vagi_suite.model_arm.preregistration.v1",
 "frozen_at": "2026-11-18T18:23:28.981742+00:00",
 "question": "Can a raw frontier model, given the same published task data and the same policy card as the VAGI Behavior Suite library arm, solve the suite's held-out tasks?",
 "not_a_claim": [
  "This does not measure CAIN's governed procedure runtime.",
  "This is not a general-intelligence benchmark.",
  "The suite is constructed; the library is hand-written for these families.",
  "No external verifier and no independent witness."
 ],
 "suite": {
  "seed": 9402,
  "heldout_tasks": 500,
  "domains": 10,
  "n_per_domain": 50,
  "checker": "vagi_suite.suite.check"
 },
 "arms": [
  {
   "id": "haiku-5-5-full",
   "model": "claude-haiku-5-5",
   "n_per_domain": 50,
   "n_tasks": 500,
   "selection": "all held-out tasks",
   "max_tokens": 400,
   "budget_cap_usd": 1.5
  },
  {
   "id": "fable-5-2-stratified",
   "model": "claude-fable-5-2",
   "n_per_domain": 10,
   "n_tasks": 100,
   "selection": "deterministic stratified sample, seed 9411, 10 per domain",
   "max_tokens": 400,
   "budget_cap_usd": 1.5
  }
 ],
 "protocol": {
  "attempts_per_task": 1,
  "invalid_json": "incorrect",
  "retries_on_wrong_answer": 0,
  "policy_card_supplied": true,
  "tools": "none",
  "memory": "none",
  "temperature": "not set (deprecated by the API)",
  "concurrency": 6
 },
 "prompt_contract": {
  "system": "one JSON object only",
  "user": "task data + domain instruction + authoritative policy card + exact JSON schema",
  "excluded_fields": [
   "fixes",
   "hidden",
   "bad",
   "reference answer",
   "other tasks"
  ]
 },
 "scoring": {
  "accepted": "vagi_suite.suite.check(domain, task, answer) exactly",
  "software_safety": "static screen rejects imports, eval/exec/open/dunder before the checker runs",
  "statistics": [
   "overall exact binomial Wilson 95% CI",
   "per-domain counts",
   "cost per accepted task",
   "latency p50/p95"
  ]
 },
 "stopping_rule": "Stop a model arm before cumulative spend reaches its budget cap. Report skipped tasks exactly. No task is retried because its answer was wrong.",
 "planned_analyses": [
  "overall and per-domain accuracy vs the published none/fixed/library/governed arms",
  "invalid-JSON rate",
  "cost and latency",
  "per-domain weak spots",
  "comparison to the 5/100 distribution-shift result (not run for the model in this pre-registration)"
 ],
 "api_pricing_usd_per_mtok": {
  "claude-haiku-5-5": {
   "input": 0.1,
   "output": 0.5
  },
  "claude-fable-5-2": {
   "input": 8.0,
   "output": 40.0
  }
 },
 "pricing_source": "https://platform.claude.com/docs/en/about-claude/pricing (read 2026-11-18)",
 "budget": {
  "requested_total_usd": 1.91,
  "hard_cap_per_arm_usd": 1.5,
  "expected": "under $1.00 for the pre-registered arms"
 }
}