1# Phase 7 — human benchmark, welfare model, robustness, interpretability (2026-10-28)23Human-appraiser arm: NOT PERFORMED (requires a UQO research-ethics certificate and recruitment); marked pending in the paper; no human numbers anywhere.4Welfare model: `paper/sections/welfare.tex`, `scripts/40_welfare_sim.py` (illustrative; propositions on rationing, lender loss, systemic correlation).5Robustness: spatial holdout (A/C/D retrained without Outaouais, Abitibi-Témiscamingue, Bas-Saint-Laurent), provider committees, rounds grid, aggregation rules, small-tier models, stratified errors (`results/evaluation/strata.csv`), run-to-run repeat, exclusion sensitivity NOT run (WP 22 exclusion flags applied as-is).6Interpretability audit: `results/evaluation/faithfulness*.csv` — cited ids 100 % valid; adjustments within ±40 % of raw comparable prices 96–99 %; point inside the adjusted-comparable hull ≈ 99 %; anchoring index 0.7–0.8; convergence 0.55 at round 2.7Fundamental-value triangulation: `scripts/50–52`, `paper/sections/fundamental.tex` (resale-implied values for 92,089 sales; rents and user cost illustrative; LLM-arm comparison deferred for lack of resale-linked subsample properties).8