# Preregistration (draft v0.1, 2026-10-27 — to be frozen before any test-set evaluation; frozen copy = results/freeze.json)

# Hypotheses (primary family, Holm-corrected, α = 0.05)

H1 (RQ1): G2 (12-role structured debate, 2 rounds, mixed committee) has lower mean absolute log error than D (AVM) on the test subsample. Two-sided. H2 (RQ2a, deliberation vs ensembling): G2 < F (same roles, no communication), paired on the same properties. H3 (RQ2b, deliberation vs compute): G2 < J (cost-matched single agent), paired. H4 (RQ3, role diversity): G2 < I (placebo roles: generic prompts with the same least-privilege evidence views and the same mechanics), paired; and G2 has lower P95 absolute log error than I. Per Proposition 9 of the theory section this contrast identifies prompt × evidence complementarity, not the value of the evidence partition itself; the partition's value is identified by the secondary contrast G2 vs G2-fullview (all twelve role prompts given the full packet). H5 (RQ3, calibration): the 90 % interval of G2 (median-of-quantiles aggregation) has coverage closer to 0.90 than E and than F. H6 (RQ4, faithfulness): ≥ 90 % of evidence ids cited by G2 agents exist in their packet, and the adjusted-comparable consistency check passes for ≥ 80 % of turns (descriptive thresholds; no test).

Secondary (exploratory, reported with CIs, no claims of significance): rounds grid 0/1/2/4; H vs G2; provider committees; stratified errors; spatial holdout; cost-adjusted Pareto frontier; repeat-sales triangulation.

# Estimands and metrics

For property $i$ with sale price $P_i$ and arm prediction $\hat P_i$: $e_i = \log \hat P_i - \log P_i$. Primary: $\bar{|e|}$. Secondary: MAE in CAD, median $|\hat P/P - 1|$, RMSE of $e$, 90th/95th percentiles of $|e|$, coverage of nominal 50/80/90/95 % intervals, mean width (log), pinball loss averaged over the five quantiles, severe mispricing rate $|e| > \log 1.25$, cost USD and wall-clock latency per property. Paired contrasts: difference in $|e_i|$ between two arms, block bootstrap with blocks = municipality × calendar month (2,000 resamples), percentile CI.

# Sample

LLM arms: stratified random subsample (region group {Montréal CMA, Québec CMA, Outaouais, other} × class {house, condo, plex} × price tercile) of the test partition, n = 600, drawn once with seed 20261027 (drawn before the freeze, never inspected; data/processed/llm_subsamples.parquet). Amendment before the freeze (2026-10-27, budget): this version of the paper evaluates the first 80 properties of the test subsample (in the fixed draw order) at the mid tier for the six main arms, the first 300 at the small cost-matched tier for E/F/G2, and the first 60 of the holdout subsample; the 600/300 draws remain the target and the runs are resumable nested extensions. Baselines: full test partition and the same subsamples (both reported). Holdout 2026 evaluated once after the test results are written.

# Decisions fixed before the freeze

Aggregation weights (inverse-variance, stacking) estimated on the validation subsample only. Prompts, roles, rounds, model ids, effort levels and packet construction are fixed by the freeze hash; any change afterwards is reported as a protocol deviation.

# Exclusions

Sales flagged by WP 22 (non-residential, land, duplicates, assessment-ratio outliers, built after sale, shared/multi-parcel) are excluded from all partitions; sensitivity in Appendix.

# What would falsify the "deliberation adds information" claim

H2 and H3 both not rejected (G2 no better than F and no better than J) with CIs excluding a 2 % relative improvement in mean absolute log error.