1% Author: Simon-Pierre Boucher — simon-pierre.boucher@uqo.ca2\section{Institutional setting: valuation and transaction records in Quebec}3\label{sec:setting}45Quebec offers an unusually clean environment for studying valuation because three public records6describe the same dwelling from three different angles, each produced by a distinct institution with7its own timing. The first is the land register (\emph{Registre foncier du Qu\'ebec}), in which every8transfer of ownership is published by a notary. A registered deed records the parties, the9consideration paid, the cadastral lot and the date of the act; it does not record the dwelling's10attributes, the marketing history, or whether the transfer was at arm's length. The second is the11municipal assessment roll, prepared under the \emph{Loi sur la fiscalit\'e municipale} by municipal12or regional assessors and supervised by the Minist\`ere des Affaires municipales et de l'Habitation13(MAMH). The roll is triennial: each unit of assessment carries a land value and a building value that14are meant to reflect market value at a reference date set eighteen months before the roll comes into15force, together with the physical attributes used by the assessor (building type, construction year,16floor area, lot area, number of storeys and dwellings, and the land-use code CUBF). The third is the17brokerage record (the MLS/Centris system operated under the authority of the OACIQ), which contains18listing prices, descriptions, photographs and marketing dates, but only for the roughly three quarters19of transactions that go through a licensed broker; it is proprietary and is not used in the arms of20this study.2122Valuation itself is a regulated profession: chartered appraisers (\'evaluateurs agr\'e\'es) are23members of the Ordre des \'evaluateurs agr\'e\'es du Qu\'ebec and produce appraisal reports for24lenders, courts and expropriations following the sales-comparison, cost and income approaches. The25agents studied here are simulated analytic roles modelled on this division of labour; none of them is26a chartered appraiser and nothing in the paper should be read as an appraisal within the meaning of27Quebec law.2829Two institutional facts shape the identification strategy. First, the assessment roll is public,30stale by construction (its reference date precedes the sale by eighteen months to four and a half31years), and known to every market participant: it is a natural, information-set-consistent anchor32that any valuation method, human or artificial, legitimately observes. We therefore give it to every33arm, together with its reference date, and we treat the price-to-assessment ratio of recent nearby34sales as the canonical ``local market state'' signal. Second, registration in the land register lags35the signature of the deed by a few days, and publication in the data platform we use lags registration36by two to eight weeks \citep[measured in a companion paper,][]{boucher2026sharp}; a valuation dated at37the deed therefore cannot use sales registered after it, and the comparable pool of every arm is38truncated at the valuation date on the \emph{sale} date, not the publication date, which is the39conservative choice.4041The sample period, January 2021 to October 2026, spans a complete cycle: the pandemic boom of422021--2022, the correction that followed the rise of policy rates in 2022--2023, and the recovery of432024--2026. This matters for the question at hand because a valuation method that merely44extrapolates recent comparables performs differently at turning points than one that reasons about45the state of the market, and because distribution shift between the training window and the test46window is a feature of the design rather than a nuisance.47