# Phase 1 — literature and novelty map (2026-10-27)

Artefacts: paper/references.bib (136 entries: 93 verified by Crossref DOI, 31 by arXiv id, 12 by live URL; 74 further verified entries held in reserve in docs/_bib_verification.json), docs/literature_map.md (themes, closest-prior-work table, RQ mapping, novelty statement, verification log), paper/sections/literature.tex (≈ 3.7 pages).

Closest prior work: Geerts et al. (2025, ECML-PKDD) single-LLM in-context valuation on four international datasets; Faxvaag et al. (2025, JRER) GPT-extracted features in an Oslo AVM; Vedula et al. (2025) fine-tuned LLM quantile regression; Chen & Si (2024) ChatGPT vs ML for Shanghai rents; Li (2026, ERES) GPT-4o with user-cost prompts vs UK self-valuations. Multi-agent debate evidence (Estornell & Liu 2024; Smit et al. 2024; Wang et al. 2024; Kaesberg et al. 2025; Zhu et al. 2026) shows debate rarely beats voting at matched cost and converges to shared misconceptions with homogeneous agents. No prior paper deploys role-specialised multi-agent LLM deliberation on residential transaction data, and none uses Quebec or Canadian data (searched Crossref, OpenAlex, Semantic Scholar, arXiv, web).

Failed checks / not found: Kahn–Ohls appraisal paper not locatable; Perry et al. 2021 update not verified (2018 report kept); Semantic Scholar rate-limited (coverage completed via OpenAlex/arXiv).

Decision: the paper claims novelty of (i) the role-partitioned deliberation design with cost- and information-matched controls on transaction data, (ii) the information-aggregation theory mapped to preregistered contrasts, (iii) the Quebec evidence. It does not claim priority on LLM valuation per se.