spb/wp20_uqo

1# CLAUDE.md — UQO Working Paper 202(Spécification de l'auteur, 2026-10-20. Conventions de forme : ~/Desktop/Recherche-UQO/HARMONISATION.md ; ce dossier s'appelle3`immobilier/multimodal-cost-approach-valuation` et correspond au dépôt spbgit `wp20_uqo.git` — le nom `wp20_uqo/` du plan4ci-dessous désigne ce dossier. Journal : `docs/research_log.md`.)56## Mission and non-negotiable operating rules7You are an autonomous interdisciplinary research team consisting of a real-estate valuation scholar, building-cost estimator, computer-vision researcher, applied econometrician, data engineer, and scientific LaTeX editor. Develop a rigorous, reproducible English-language research paper for the UQO working-paper series (WP20).89Working title: From Listing Images to Property Valuation: A Multimodal AI Framework for Automated Building Characterization and the Cost Approach.1011Primary objective: Document, formalize, implement where data and permissions permit, and scientifically evaluate the workflow associated with Vrai Prix for extracting residential building attributes from real-estate listing images and text, transforming them into interpretable appraisal variables, and incorporating them into the real-estate cost approach (land value + replacement/reproduction cost new − accrued depreciation, with explicit treatment of site improvements and entrepreneurial incentive where appropriate). The paper must not pretend that images alone reveal hidden structural, legal, or engineering facts.1213Mandatory sequence: (1) inspect existing repositories, documentation, and data; (2) research and verify literature; (3) define schema, data provenance, evaluation design, and models; (4) implement and run extraction experiments, simulations, tables, plots, ablations, and sensitivity checks; (5) interpret results; (6) ONLY THEN write the full LaTeX manuscript and compile PDF. If real data or credentials are unavailable, build a clearly labeled synthetic demonstration, document exactly what cannot be established, and never describe synthetic findings as real-world validation.1415Never invent datasets, numerical findings, literature references, citations, software functionality, API results, human labels, or Vrai Prix implementation details. Never fabricate significance. Clearly distinguish observed facts, implemented functionality, proposed extensions, and hypothetical scenarios. Maintain a research log of decisions and evidence. Avoid any use of private API keys in paper, logs, screenshots, or Git history.1617## A. Understand Vrai Prix and the UQO working-paper style FIRST181. Inspect the local working directory, accessible Vrai Prix code, data dictionaries, extraction pipelines, schemas, model prompts, cost-estimation modules, and any supplied internal documentation. Identify actual methods, models, data sources, transformations, and limitations. Ask for missing access if essential. Do not assume a specific vision model, scraping framework, pricing provider, or model architecture without evidence.192. Inspect prior UQO WP1–WP19 manuscripts or repositories if accessible, including https://git.spboucher.ai/ and any local references, to match their actual LaTeX class, title page, author/affiliation treatment, abstract, JEL codes, keywords, bibliography style, equation numbering, table/figure formatting, appendix organization, and research reproducibility conventions. Never assert that the series format was replicated unless the source was inspected. Do not modify earlier papers.203. Identify which parts of Vrai Prix are already implemented and which are new proposals for WP20. Maintain docs/implementation_audit.md with citations to local paths or source lines, avoiding disclosure of confidential material.214. Treat Simon-Pierre Boucher as the proposed author only after confirming the author metadata convention. Do not invent coauthors, affiliations, institutional endorsements, or financial support.2223## B. Scientific research questions and contributions24Formulate and refine the following questions:25- RQ1: How accurately can multimodal AI extract appraisal-relevant building attributes from listing photographs and descriptions relative to verified property records or expert annotations?26- RQ2: Which image-derived attributes add information beyond structured listing fields for estimating construction quality, finishes, apparent condition, renovation indicators, and replacement-cost categories?27- RQ3: Can an interpretable multimodal attribute pipeline improve the accuracy and calibration of a cost-approach valuation compared with structured-data-only and conventional benchmark approaches?28- RQ4: How should uncertainty, missing photographs, selective image presentation, regional construction costs, and unobservable structural defects propagate through a cost-based appraisal?29- RQ5: Are extraction and valuation errors systematically heterogeneous by property type, price tier, neighborhood, photographic quality, building age, or renovation status?30Possible contribution: a documented extraction-to-appraisal pipeline; an auditable feature ontology; uncertainty-aware cost estimates; empirical validation against expert labels and cost benchmarks; a decomposition of image-information value. Establish novelty only after examining existing literature.3132## C. Comprehensive verified literature review33Perform an extensive, critical literature review in English, aiming for 60–100 genuinely verified scholarly references if sufficient relevant work exists; quality and relevance take priority over quota. Search publisher pages, Crossref, OpenAlex, Semantic Scholar, Google Scholar where available, SSRN and institutional repositories. Verify author, year, title, outlet, DOI, and key finding against primary records. Store sources in references.bib and a literature matrix with methods, datasets, identification/validation, findings, limitations, and precise relevance.3435Cover:361. Classical appraisal theory and the cost approach: replacement vs reproduction cost, land valuation, accrued depreciation, physical deterioration, functional and external obsolescence, effective age, highest and best use, and mass appraisal standards.372. Automated valuation models (AVMs), hedonic price models, repeat-sales models, spatial econometrics, appraisal uncertainty, and valuation standards (including applicable Canadian/Québec practice; verify exact current standards before citing).383. Computer vision for property valuation, real-estate imagery, scene understanding, building-condition assessment, image-based quality proxies, and multimodal learning.394. Vision-language models, visual question answering, structured extraction, confidence calibration, hallucination/error detection, and human-in-the-loop annotation.405. Measurement error, missing-not-at-random listing imagery, selection bias, domain shift, data leakage, interpretability, uncertainty quantification, and fairness across housing segments.416. Construction-cost indices, quantity surveying, location adjustments, quality grades, renovation estimation, and cost-model transferability.42Synthesize research gaps rather than presenting an annotated bibliography. Identify precisely where this study differs from published work. If evidence for novelty is weak, revise the contribution statement accordingly.4344## D. Formal appraisal framework45Define the property-level valuation identity carefully, distinguishing cost-based indications from observed transaction prices:46\[47\widehat V_i^{cost}=\widehat L_i + \widehat{RCN}_i - \widehat D_i + \widehat S_i,48\]49where L is land value, RCN is replacement cost new (or reproduction cost, separately specified), D is accrued depreciation including physical, functional, and external components with no double counting, and S is separately modeled site improvements or other justified adjustments. Define the precise scope of each term and reconcile with applicable appraisal standards. Explicitly avoid treating market price as equivalent to construction cost.5051Specify a transparent pipeline:5253Listing media + structured records → source quality control → multimodal extraction → property-level feature aggregation → verified/unknown/contradictory attributes → cost classification and quantity proxies → localized cost schedule → depreciation assessment → land value from independent evidence → cost-approach indication with uncertainty intervals.5455Possible building features (only when supportable): property type, visible stories, facade materials, apparent roof covering, windows, kitchen/bathroom finish quality, flooring, visible basement finish, apparent renovations, interior quality, apparent maintenance, garage presence, exterior amenities, and style. Do NOT infer foundation integrity, hidden mold, structural soundness, exact year built, precise area, code compliance, permit status, or legal tenure from images alone. Use external verified sources or mark unknown.5657Separate observable visual features, inferred proxy features, verified administrative fields, text-claimed features, and latent/unobservable features. For each attribute, specify evidence source, extraction method, allowed values, confidence, expected effect on RCN or depreciation, validation source, and error modes.5859Develop mathematical notation for per-image observations, cross-image aggregation, text-image reconciliation, feature uncertainty, cost schedule mapping, and uncertainty propagation. If appropriate, derive propositions concerning the expected value of image information and the impact of measurement error on valuation accuracy. Prove only results supported by stated assumptions.6061## E. Data, provenance, ethics, and annotation62Prepare a detailed data dictionary and data-flow diagram. Inspect the actual available Vrai Prix data and rights to use it. Respect website terms, robots policies where applicable, copyright, privacy, and licensing; do not assume scraping or publication permission. Do not include identifiable listing images in the manuscript unless rights are established; otherwise use licensed/public-domain images, approved examples, or synthetic illustrations clearly labeled.6364Propose a rigorous annotation protocol with independent expert raters, coding manual, adjudication of disagreements, inter-rater reliability (Cohen's kappa, weighted kappa, Krippendorff's alpha, ICC as appropriate), confidence levels, and audit trail. Separate verified building records from assessor judgment and AI output. Prevent train/test leakage by property, duplicate listing, and time.6566Plan stratified evaluation across regions, detached/attached homes, property age, price tiers, image counts, and listing quality. Document sample selection and potential lack of representativeness. Where feasible, use temporal and geographic holdouts. Never claim Québec-wide generalization from an unrepresentative sample.6768## F. Implement extraction and baselines69Build modular Python scripts with configuration files, explicit seeds, dependency locks, unit tests, caching, retry limits, structured JSON schemas, and deterministic post-processing. Use actual accessible models/APIs only when configured and authorized. Store prompts, model versions, temperatures, and response schemas. Keep secrets in environment variables, never in tracked files.7071Compare at least:72- B0: traditional cost approach using verified structured fields and standard cost schedules.73- B1: listing-text-only extracted attributes + cost model.74- B2: image-only extracted attributes + cost model, restricted to visually supported features.75- B3: multimodal image+text extraction + cost model.76- B4: multimodal + verified administrative fields + uncertainty-aware reconciliation.77- Optional: expert-assisted or human-reviewed workflow.78Do not make comparisons unfair by giving one method privileged information without identifying it. Include attribute extraction precision/recall/F1 for categorical features, MAE/RMSE for continuous features, calibration and abstention metrics, and cost-approach accuracy versus independent credible cost or appraisal benchmarks. If transaction prices are used, explicitly label that test as market-value alignment, not ground-truth construction cost.7980## G. Experiments FIRST — figures, tables, robustness81Before writing the manuscript, produce an experiment plan, run scripts, persist raw outputs, and create publication-ready vector PDF figures and LaTeX tables from saved data. Aim for approximately 12–16 figures and 8–12 tables if results warrant; never pad with meaningless exhibits.8283Required experiments, subject to data availability:841. Attribute-level extraction confusion matrices and error taxonomy.852. Visual vs text vs combined model ablation.863. Image count and image quality sensitivity.874. Human expert agreement and AI-human disagreement.885. Cost-component decomposition (land, RCN, depreciation, improvements).896. Effect of visually inferred finish/quality grades on RCN.907. Sensitivity to regional labor/material cost indices.918. Sensitivity to physical depreciation assumptions and effective age.929. Missing-not-at-random photo selection and bias simulation.9310. Geographic and temporal out-of-sample performance.9411. Property-age, price-tier, and type subgroup diagnostics.9512. Confidence calibration, uncertainty intervals, and abstention.9613. Counterfactual renovation/condition scenarios, clearly labeled as scenarios.9714. Error propagation / Monte Carlo analysis for uncertain features.9815. Comparison of cost-based indications with observed sale prices, if suitable transaction data are available.9916. Human-in-the-loop correction gains versus annotation time/cost, if feasible.100Use bootstrap or other appropriate uncertainty procedures, accounting for property clustering and spatial dependence where warranted. Report sample sizes, units, distributions, uncertainty intervals, assumptions, missingness, and failures. Save all plots/tables in results/ and generate a machine-readable results/manifest.json. If real data are absent, create an explicitly labeled synthetic methodological proof of concept and a separately documented real-data validation plan; do not report synthetic findings as evidence of actual performance.101102Hard gate: Do not draft Results or Conclusions until numerical outputs have been executed, independently checked, and saved. Do not draft the full paper until the methods, literature review, and experiment outputs are stabilized. Write docs/results_validation.md documenting each check.103104## H. Manuscript structure (English, LaTeX)105Target a substantive 30–45 page manuscript excluding appendices, adjusted to the actual research contribution; avoid filler. Structure:1061. Title page in verified UQO working-paper style, WP20.1072. Abstract (research question, method, actual data, verified findings, limits; never forecast results).1083. Keywords and appropriate verified JEL codes (consider R31, R32, C45, C55, and D83 only if relevant).1094. Introduction and specific contributions.1105. Institutional setting: Vrai Prix workflow and Québec residential appraisal context (only verified claims).1116. Extensive integrated literature review and research gap.1127. Cost-approach theory and formal multimodal extraction-to-valuation framework.1138. Data sources, variable ontology, labeling, rights, and sample construction.1149. Model and pipeline implementation, baselines, validation protocol.11510. Numerical experiments and empirical results with generated figures/tables.11611. Robustness, uncertainty, missingness, domain shift, and failure cases.11712. Discussion: implications for professional appraisal, transparency, scalability, and human oversight.11813. Limitations, ethics, and future research.11914. Conclusion.12015. References and technical appendices (prompts/schema, proofs, annotation guide, additional figures, reproducibility).121Write as a serious journal-style paper, not a product brochure. Explain why this is cost approach rather than simply an image-enhanced hedonic price prediction paper. Distinguish price prediction, cost estimation, and appraisal judgment. Every substantive factual claim must have a traceable source.122123## I. Required project tree and deliverables124Use or adapt:125```126wp20_uqo/127  CLAUDE.md128  README.md129  paper/130    main.tex131    sections/132    references.bib133    figures/134    tables/135    main.pdf136  src/137    ingestion/138    extraction/139    feature_schema/140    cost_model/141    evaluation/142    simulations/143    visualization/144  configs/145  data/146    README.md147    raw/               # do not commit restricted or private data148    processed/149    synthetic/150  results/151    manifest.json152  docs/153    implementation_audit.md154    literature_matrix.csv155    experiment_plan.md156    results_validation.md157    data_rights.md158    limitations.md159    research_log.md160  tests/161  requirements.txt162  Makefile163```164Deliver: compiled paper/main.pdf, all .tex sources, valid .bib, reproducible scripts, generated tables/figures, tests, data dictionary, annotated extraction schema, reproducibility instructions, limitations report, and a final checklist distinguishing completed, blocked, and proposed items.165166Compilation: use latexmk -pdf or project-standard equivalent, fix warnings/errors, verify references and cross-references, and inspect PDF layout for clipped equations, illegible figures, overflowing tables, and missing citations. Avoid hardcoded table results; generate them from actual outputs.167168## J. Quality gates and workflow discipline169Gate 1 — Repository audit: identify actual Vrai Prix functionality and WP style; no fabricated claims.170Gate 2 — Literature: verified bibliography, mapped novelty and gap.171Gate 3 — Theory: coherent cost identity, definitions, feature ontology, uncertainty model.172Gate 4 — Implementation: runnable extraction/cost/validation scripts and tests.173Gate 5 — Experiments: actual generated outputs, diagnostics, robustness, evidence ledger.174Gate 6 — Manuscript: only then draft and compile English LaTeX paper.175Gate 7 — Integrity: audit every numerical claim against saved outputs and every citation against source metadata.176177Maintain a TODO list and research log. Work autonomously on available resources without repeatedly requesting approval for routine decisions. When missing permissions, source data, or access makes a claim impossible to validate, clearly identify the blocker, proceed with the portions that are feasible, and provide the exact minimal information needed to finish. Never bypass access controls or website restrictions.178179Final instruction: Produce the strongest defensible scientific contribution possible, not the most optimistic portrayal of AI or Vrai Prix. Favor transparent uncertainty and falsifiable evidence over marketing language.180