# API-cost tiers. Costs are ESTIMATES from configs/models.yaml list prices and the token counts observed in the
# live sanity run of 2026-10-27 (single-agent turn at the mid tier: ~5.5k input + ~1.0k output tokens ≈ $0.02;
# F ≈ 12 turns ≈ $0.25; G2 ≈ 60 turns ≈ $1.2-1.4 per property). Actual spend is logged per call in results/ledger/.
# Self-imposed envelope for this version of the paper: ≈ 1,000 USD in total. Every run is resumable, so any tier can
# be extended later by raising `n` (the subsample order is fixed by the seed, so extensions are nested samples).
tier_default: mid
tiers:
  pilot:
    split: validation
    n: 12
    runs:
      - {arm: E_single_llm, committees: [all_anthropic, all_openai]}
      - {arm: F_independent, committees: [all_anthropic, all_openai]}
      - {arm: G_debate, rounds: 2, committees: [all_anthropic, all_openai]}
    cap_usd: 60
    purpose: "smoke on real validation data; prompt bugs; cost calibration"
  validation:
    split: validation
    n: 40
    runs:
      - {arm: E_single_llm, committees: [all_anthropic, all_openai, mixed_random]}
      - {arm: F_independent, committees: [all_anthropic, all_openai, mixed_random]}
      - {arm: G_debate, rounds: 2, committees: [all_anthropic, all_openai, mixed_random]}
      - {arm: G_debate, rounds: 1, committees: [mixed_random]}
      - {arm: G_debate, rounds: 4, committees: [mixed_random], n: 20}
      - {arm: H_debate_skeptic, rounds: 2, committees: [mixed_random]}
      - {arm: I_placebo, rounds: 2, committees: [mixed_random]}
      - {arm: J_cost_matched, committees: [mixed_random]}
      - {arm: G_fullview, rounds: 2, committees: [mixed_random]}
    cap_usd: 500
    purpose: "ablations, aggregation-weight estimation, protocol freeze"
  final:
    split: test
    n: 80
    runs:
      - {arm: E_single_llm, committees: [mixed_random]}
      - {arm: F_independent, committees: [mixed_random]}
      - {arm: G_debate, rounds: 2, committees: [mixed_random]}
      - {arm: H_debate_skeptic, rounds: 2, committees: [mixed_random]}
      - {arm: I_placebo, rounds: 2, committees: [mixed_random]}
      - {arm: J_cost_matched, committees: [mixed_random]}
    cap_usd: 500
    purpose: "test set, evaluated once after the freeze"
  final_small:
    split: test
    n: 300
    tier: small
    runs:
      - {arm: E_single_llm, committees: [mixed_random]}
      - {arm: F_independent, committees: [mixed_random]}
      - {arm: G_debate, rounds: 2, committees: [mixed_random]}
    cap_usd: 120
    purpose: "same protocol at the small cost-matched tier (claude-haiku-5-5 / gpt-6-luna): power + cost frontier"
  holdout:
    split: holdout
    n: 60
    runs:
      - {arm: E_single_llm, committees: [mixed_random]}
      - {arm: F_independent, committees: [mixed_random]}
      - {arm: G_debate, rounds: 2, committees: [mixed_random]}
    cap_usd: 120
    purpose: "untouched 2026 holdout, evaluated once"
scale_up_estimates_usd:
  note: "list-price estimates for extending the test tier (mid tier, mixed committee, six arms): n=200 ≈ 1,100; n=600 ≈ 3,300"