# Phase 2 — preregistration, estimands, theory, data governance (2026-10-27)

Artefacts: docs/preregistration.md (H1–H6, primary family with Holm, estimands, sample, decisions fixed before the freeze; H4 amended after the theory work: the placebo contrast identifies prompt × evidence complementarity, the evidence-partition value is identified by G2 vs G2-fullview), paper/sections/theory.tex (11 propositions + lemma, proofs in paper/appendix/proofs.tex), src/agentic_valuation/theory/ + scripts/41_theory_sim.py (Monte Carlo checks: closed forms match simulation, 19 tests), paper/sections/welfare.tex + scripts/40_welfare_sim.py (illustrative buyer–seller–lender model; two figures), docs/data_dictionary.md, docs/data_access_checklist.md.

Key theoretical results used in the design: the equal-weight pool has MSE floor ρσ² under correlated signals (K cannot remove it); opinion-only revision toward the committee median contracts dispersion by (1−λ)^r without reducing MSE (and slightly raises it with a median anchor); evidence exchange reduces MSE only for evidence outside the common view; any linear deliberation cannot beat the optimal round-0 pool; permuting prompts with fixed views is distributionally vacuous under additive separability.

Outstanding assumptions: linear-Gaussian signals, known covariance for optimal weights; the welfare model is illustrative (LTV 80 %, haircut rule) and is not calibrated to Quebec lending data.