1# API-cost tiers. Costs are ESTIMATES from configs/models.yaml list prices and the token counts observed in the2# live sanity run of 2026-10-27 (single-agent turn at the mid tier: ~5.5k input + ~1.0k output tokens ≈ $0.02;3# F ≈ 12 turns ≈ $0.25; G2 ≈ 60 turns ≈ $1.2-1.4 per property). Actual spend is logged per call in results/ledger/.4# Self-imposed envelope for this version of the paper: ≈ 1,000 USD in total. Every run is resumable, so any tier can5# be extended later by raising `n` (the subsample order is fixed by the seed, so extensions are nested samples).6tier_default: mid7tiers:8 pilot:9 split: validation10 n: 1211 runs:12 - {arm: E_single_llm, committees: [all_anthropic, all_openai]}13 - {arm: F_independent, committees: [all_anthropic, all_openai]}14 - {arm: G_debate, rounds: 2, committees: [all_anthropic, all_openai]}15 cap_usd: 6016 purpose: "smoke on real validation data; prompt bugs; cost calibration"17 validation:18 split: validation19 n: 4020 runs:21 - {arm: E_single_llm, committees: [all_anthropic, all_openai, mixed_random]}22 - {arm: F_independent, committees: [all_anthropic, all_openai, mixed_random]}23 - {arm: G_debate, rounds: 2, committees: [all_anthropic, all_openai, mixed_random]}24 - {arm: G_debate, rounds: 1, committees: [mixed_random]}25 - {arm: G_debate, rounds: 4, committees: [mixed_random], n: 20}26 - {arm: H_debate_skeptic, rounds: 2, committees: [mixed_random]}27 - {arm: I_placebo, rounds: 2, committees: [mixed_random]}28 - {arm: J_cost_matched, committees: [mixed_random]}29 - {arm: G_fullview, rounds: 2, committees: [mixed_random]}30 cap_usd: 50031 purpose: "ablations, aggregation-weight estimation, protocol freeze"32 final:33 split: test34 n: 8035 runs:36 - {arm: E_single_llm, committees: [mixed_random]}37 - {arm: F_independent, committees: [mixed_random]}38 - {arm: G_debate, rounds: 2, committees: [mixed_random]}39 - {arm: H_debate_skeptic, rounds: 2, committees: [mixed_random]}40 - {arm: I_placebo, rounds: 2, committees: [mixed_random]}41 - {arm: J_cost_matched, committees: [mixed_random]}42 cap_usd: 50043 purpose: "test set, evaluated once after the freeze"44 final_small:45 split: test46 n: 30047 tier: small48 runs:49 - {arm: E_single_llm, committees: [mixed_random]}50 - {arm: F_independent, committees: [mixed_random]}51 - {arm: G_debate, rounds: 2, committees: [mixed_random]}52 cap_usd: 12053 purpose: "same protocol at the small cost-matched tier (claude-haiku-5-5 / gpt-6-luna): power + cost frontier"54 holdout:55 split: holdout56 n: 6057 runs:58 - {arm: E_single_llm, committees: [mixed_random]}59 - {arm: F_independent, committees: [mixed_random]}60 - {arm: G_debate, rounds: 2, committees: [mixed_random]}61 cap_usd: 12062 purpose: "untouched 2026 holdout, evaluated once"63scale_up_estimates_usd:64 note: "list-price estimates for extending the test tier (mid tier, mixed committee, six arms): n=200 ≈ 1,100; n=600 ≈ 3,300"65