% ============================================================================
% Conclusion (WP 26)
% Author: Simon-Pierre Boucher — simon-pierre.boucher@uqo.ca
% ============================================================================
\section{Conclusion}
\label{sec:conclusion}

This paper asked whether communication among heterogeneous autonomous valuation agents adds measurable information about the value of a dwelling, and it answered with a preregistered experiment on registered Quebec transactions in which every rival explanation for a deliberation effect was given its own control arm. The answer, on the evidence of this version of the study, is no. Twelve role-specialised agents that read least-privilege views of the same timestamped evidence packet and never communicate reach the same accuracy as the same agents after two rounds of critique and revision, as a committee with the mandates removed, and, within a wider interval, as one agent given the same tokens; the four preregistered contrasts have point estimates within 0.006 log points of zero and Holm-adjusted $p$-values of one, and in the 2026 holdout and the prospective set the debated committee is marginally worse than the independent one. What deliberation does, reliably and in every run, is manufacture agreement: the dispersion of the agents' estimates halves in the first round and keeps shrinking thereafter while the error of their median does not move, which is what the theory of Section~\ref{sec:theory} says must happen when agents revise toward a committee statistic rather than toward new evidence. The committees themselves are competent price forecasters, indistinguishable from a tuned automated valuation model on the test subsample and better on the median error, but they state their uncertainty too narrowly, their small-model versions markedly so, and when the market moved in 2026 all of them moved with it in the same direction, overvaluing by ten log points where the tabular model was unbiased. None of this is evidence that the agents have discovered fundamental value. By Lemma~\ref{lem:proxy} the price-based metrics rank estimators by their distance to the selected market value, and the only proxy for value that is independent of the first transaction, the market-adjusted resale, is not yet available for enough of the evaluated properties to say anything about the agentic arms.

For lenders, appraisers and the providers of valuation models the findings carry three practical messages. The first concerns common-model risk. A committee whose members are drawn from two foundation models inherits the errors those models share, and Proposition~\ref{prop:nested} shows that no committee size removes them; the 2026 drift is a direct illustration, since every agent, whatever its role or provider, overvalued the same properties in the same direction. A lending system in which every participant consults the same committee therefore gains little in expected accuracy over a good tabular model and loses the diversification that independent valuation errors provide, and Proposition~\ref{prop:systemic} shows that the cost appears in the dispersion of aggregate losses rather than in their mean. The second concerns verification. The deterministic checks of arm~H, which flag citations to evidence not in the packet, adjustments far from the raw comparable price and points outside the hull of the comparables, cost almost nothing, catch the failures that would be most embarrassing in a lending file, and have no measurable effect on accuracy; they are a governance tool, not a performance tool, and should be adopted as such. The third concerns calibration. The intervals that agents state are narrower than the evidence warrants and the agreement of a committee after deliberation is not a measure of confidence; a lender that used either as its haircut rule would be systematically under-protected, and a split-conformal rescaling estimated on held-out data, which the paper reports as a remedy, is the minimum that responsible use requires.

Several changes to the design could overturn the conclusion, and we state them so that the null is read as a statement about this experiment rather than about deliberation in general. The sample is small: eighty properties at the mid tier rule out relative improvements larger than about four percent against the independent committee and about seven percent against the compute-matched agent, and the preregistered six hundred, which the nested design allows to be reached by resuming the same runs, would bring the resolution to the two percent threshold that the preregistration set as the falsification criterion. The evidence is thin: the packet contains the roll attributes, twelve comparables, a market summary and a geographic context, and every role sees a subset of the same four sections, so that the union of the views is small and, by Proposition~\ref{prop:exchange}, the scope for evidence exchange to help is correspondingly small. Richer and more specialised evidence, listing text and photographs of the kind that \citet{nowak2017textual}, \citet{zhang2024describe} and \citet{faxvaag2025incorporating} show to be informative, inspection reports, permit histories, or tools that let an agent query the register rather than read a fixed excerpt, would enlarge the union of the views and could create the complementarity that the placebo contrast found absent. The agents are general-purpose models prompted once; fine-tuning on Quebec transactions in the manner of \citet{vedula2025quantile}, or a calibration layer trained on the committee's own validation errors, would address the interval problem directly and might also change the round-0 accuracy on which everything else is built. And the deliberation protocol is one of many: the anonymised table and median anchor of arm~G are the natural first design, but a protocol that forced agents to exchange evidence rather than opinions, or that weighted members by demonstrated precision as \citet{golub2010naive} requires, could in principle move the pooled error, although Proposition~\ref{prop:degroot} bounds what any linear scheme can achieve.

The human arm remains to be run. A panel of chartered appraisers valuing the same packets under the same cutoff, with prices blinded and confidence intervals recorded, is the comparison that practitioners will ask for first and the one that would settle whether the committee's accuracy is competitive with professional judgement on identical information; it requires a research-ethics certificate that has not yet been obtained, and the paper reports it as pending rather than simulating it. Until that arm exists, and until the evaluated properties have resold, the paper's claims are deliberately limited: a deliberating committee of language-model agents predicts Quebec transaction prices about as well as an automated valuation model, no better than the same agents kept silent, and with intervals that should not be trusted at face value. That is a smaller result than the question in the title invites, but it is the one the data support, and the artefacts released with the paper are intended to let it be revised rather than repeated.