Origin: « Rôle d'évaluation foncière du Québec », Ministère des Affaires municipales et de l'Habitation, published on Données
Québec under CC-BY 4.0 (CKAN metadata read 2026-10-20; docs/institutional_sources.md). Owner names, cadastre and postal
addresses are redacted at source and re-identification is prohibited; our copy (vraiprix.db → units) contains no owner data.
Use in WP20: unit characteristics and assessed land/building values as verified administrative fields; attribution given in the
paper. Civic addresses are used only to match listings and are never printed.
# 2. Transactions (Registre foncier via Vrai Prix)
Used only in aggregate (Experiment 15: ratio of recorded sale price to cost indication for ≈170 frame units sold after listing).
No individual transaction is reproduced. Vrai Prix's own policy (Toveli integration matrix) is "not republished in bulk".
Collected by the author's Immo-Ka connectors for the Vrai Prix/Immo-Ka platforms (not by this project). Descriptions are
copyrighted marketing texts of brokers/owners; WP20 processes them automatically and reports only aggregate statistics and
short anonymised fragments (≤ 15 words) in the error taxonomy. Prices, taxes and municipal assessments are masked before any
model sees the text (src/extraction/textprep.py).
Copyright belongs to the listing broker/owner/photographer. WP20 does not download, store or redistribute any
photograph: images are passed to the model API by URL and are not kept (data/images/ is git-ignored and unused); no listing
photograph appears in the manuscript. Illustrations in the paper are schematic drawings or synthetic images clearly labelled.
Robots.txt policy: hosts whose robots.txt disallows all automated retrieval (media.remax-quebec.com) were excluded from
the extraction frame (rule R10) rather than circumvented; the model provider's fetcher enforces the same rule. Other hosts
either publish no robots.txt or do not disallow image paths (results/raw/robots_check.txt). Terms of use of individual
broker sites were not reviewed one by one; this is recorded as a residual legal uncertainty, and the analysis is a
non-commercial academic study of aggregate extraction behaviour.
Privacy: photographs may show personal belongings or people; the extraction schema contains no field about occupants, and the
evidence free-text field is instructed to name building cues only. Outputs are reviewed for incidental personal information
before release (scripts/ check: no names, plates or faces are requested or stored).
Three Anthropic models via the Message Batches API; prompts, schema versions and model ids are versioned
(configs/prompts.yaml, configs/models.yaml, results/raw/batches/run_meta.json). API keys live only in an untracked file
outside the repository; results/raw/api_usage.jsonl records tokens, never content.
Model outputs are treated as AI claims, never as ground truth; they are compared with administrative and broker fields.
No human expert annotations exist; the annotation protocol (docs/annotation_protocol.md) is a proposal.
The representativeness of the frame is limited (listings from 65 connectors, Aug–Oct 2026, RE/MAX excluded by R10, Kijiji
and Facebook excluded by R8); no Québec-wide generalisation is claimed.