# Treatment arms (CLAUDE.md §Deliberation protocols). Letters are canonical and used in tables.
arms:
  A_hedonic:        {family: baseline, model: hedonic_ridge, note: "log-price OLS/ridge with municipality x quarter and grid fixed effects"}
  B_hierarchical:   {family: baseline, model: hierarchical_hedonic, note: "mixed-effects hedonic with municipality and grid random intercepts; repeat-sales adjustment when the parcel has a prior sale"}
  C_gbm:            {family: baseline, model: lightgbm, note: "LightGBM tuned on validation only; CatBoost/XGBoost reported as robustness"}
  D_avm:            {family: baseline, model: avm, note: "GBM + comparable-based features + spatial-temporal lags + conformal / quantile intervals"}
  E_single_llm:     {family: llm, roles: [generalist], rounds: 0}
  F_independent:    {family: llm, roles: default12, rounds: 0, aggregation: all}
  G_debate:         {family: llm, roles: default12, rounds: [1, 2, 4], protocol: structured_debate}
  H_debate_skeptic: {family: llm, roles: default12, rounds: 2, protocol: structured_debate, blinded_skeptic: true, evidence_verification: true}
  I_placebo:        {family: llm, roles: shuffled12, rounds: 2, protocol: structured_debate, note: "role prompts randomly permuted / generic placeholder roles"}
  G_fullview:       {family: llm, roles: default12_fullview, rounds: 2, protocol: structured_debate, note: "secondary contrast: default prompts, full packet for every role (value of the evidence partition)"}
  J_cost_matched:   {family: llm, roles: [generalist], repeats: 12, rounds: 0, note: "single agent sampled k times to match the token budget of F/G; plus self-refine chain"}
aggregation_rules: [mean, median, trimmed_mean_20, inverse_variance, stacking, robust_bayes]
rounds_grid: [0, 1, 2, 4]
providers:
  committees: [all_anthropic, all_openai, mixed_random]
  tier: mid