# Limitations report — WP20 (what cannot be established with these data)

1. **No human expert labels.** Agreement is measured against broker fields (claims) and the roll (few attributes). Finish
   grades and condition scores are validated only by cross-extractor agreement (lower bound on reliability, nothing on
   validity). The expert protocol is written (`docs/annotation_protocol.md`) but not executed.
2. **Latent attributes are unobservable** (foundation, hidden moisture, structure, permits, tenure, exact area) and are never
   inferred; Monte Carlo intervals are conditional on their absence and under-cover prices and assessor totals (≈ 28 % vs 80 %
   nominal), which is reported as a measurement of what is omitted.
3. **No cross-city construction cost levels.** StatCan 18-10-0289 gives temporal change by CMA, not levels; the two unit-cost
   anchors are a vendor's secondary reporting (Altus via NerdWallet) and builder-facing guides (Soumission Rénovation, Dessins
   Drummond). The regional level factor is a sensitivity parameter (±10 %; non-metro −15 % scenario).
4. **Assessor values are judgments** at a 2022–2024 market date, indexed with a construction price index; alignment with them
   is a benchmark, not accuracy. The roll-vintage split shows the indexing is imperfect.
5. **Asking prices are not transactions.** The recorded-sales test covers 146 + 24 self-selected quickly sold units only.
6. **Representativeness.** Frame = listings from 65 connectors over 10 weeks (2026-08-12 → 2026-10-17), matched ≤ 15 m, ≥ 4
   photos, excluding Kijiji/Facebook (no structured fields) and RE/MAX Québec (robots.txt), one listing per unit. RE/MAX
   exclusion removes 5,287 listings that are more urban with higher land values. The stratified sample over-represents small
   CMAs by design (weights stored). No claim about the Québec housing stock.
7. **One model family** (Claude Haiku/Sonnet/Opus 5.5); cross-extractor agreement understates cross-family disagreement.
8. **Vrai Prix production cost estimate** used as a context benchmark has an uninspected formula (code unavailable offline).
9. **Image budget**: 12 (Haiku) / 10 (Sonnet, Opus) photographs in seller order; listings have a median of 36. Basement and
   bathroom coverage would rise with more photographs but, as the K-curves show, slowly.
10. **Depreciation model** is straight-line age–life with a condition factor (assumption); the age gradient of the alignment
    ratios shows it is too steep for old houses and too lenient for new ones. Functional and external obsolescence are zero in
    the baseline (scenario 10 % external).
11. **Rights**: media processed by URL only, not stored; robots-disallowing hosts excluded; site terms of use not reviewed
    one by one (residual legal uncertainty, `docs/data_rights.md`).
12. **Blocked during the study**: cluster/gateway unreachable (laptop off LAN) → Vrai Prix production repository and `cost.db`
    not re-inspected; spbgit push deferred; MEFQ Part 3E text not retrieved; OEAQ standards edition not confirmed.