# UPGRADE_REPORT — scholarly upgrade of the paper (2026-08-05)

Scope: literature and framing only. No result, table value, figure, or data statistic was changed. Paper: 47 → 53 pages; references: 59 → 86, every entry now verified. Compiles clean (0 errors, 0 undefined citations, 0 BibTeX warnings). Companion documents: PAPER_REVIEW.md (pre-upgrade critique).

# 1. ⚠️ Hallucinated citations found in the ORIGINAL bibliography

While verifying, I audited the 18 most load-bearing existing entries. 8 were unverifiable and almost certainly fabricated — searches of OpenAlex, Crossref, and the web found no such papers (in several cases the claimed journal issue/pages exist and contain something else):

4 more had wrong metadata, now corrected (same keys kept): nowak2017quality (true venue: Journal of Applied Econometrics 32(4), not JREFE), shen2020text (actually Shen & Ross 2021, Journal of Urban Economics 121), gibbons2014costs (actually Gibbons 2015, JEEM 72, sole-authored), cheshire2004capitalisation (true title: "Capitalising the Value of Free Schools", Economic J.). Six others were verified as-is and completed with full metadata + DOI (kok2017big, marinescu2020opening, haurin2010list, yoo2012variable, wang2020minilm, tyrvainen2005benefits). One never-cited, unverified entry (pace1998appraisal) was deleted.

Every sentence that cited a removed key was rewritten around its verified replacement — claims were adjusted to what the real papers actually show.

# 2. New references added (28, all OpenAlex-verified with DOI)

Text-as-data in economics (new positioning axis):

• Gentzkow, Kelly & Taddy 2019, JEL — canonical survey; frames our method as a semantic dictionary.
• Ash & Hansen 2023, Annu. Rev. Econ. — deep-learning-era survey for economists.
• Hansen, McMahon & Prat 2017, QJE — flagship interpretable-text application (FOMC).
• Loughran & McDonald 2016, JAR — dictionary-method survey; our approach's direct ancestor.
• Varian 2014, JEP; Dell 2025, JEL — ML-for-econometrics practice anchors.

Hedonic origins & identification:

• Griliches 1971 (Harvard UP) — hedonic index origins (Court 1939 not indexed anywhere; mentioned in prose only).
• Ekeland, Heckman & Nesheim 2004, JPE; Bajari & Benkard 2005, JPE; Kuminoff, Smith & Timmins 2013, JEL — grounds the "equilibrium gradients, not WTP" stance and the sorting interpretation of negative coefficients.

Quantile hedonics (previously uncited despite §7.3):

• Zietz, Zietz & Sirmans 2007, JREFE; McMillen 2008, JUE — our Luxury-gradient finding now explicitly agrees with both.

Listing language & disclosure (closest competing literature, previously absent):

• Haag, Rutherford & Thomson 2000, JRER; Pryce & Oates 2008, Housing Studies; Grossman 1981, JLE (pairs with Milgrom 1981).

Spatial hedonics & Quebec:

• Anselin 1988 (book); Dubin 1988, REStat; Can 1992, RSUE; Dubé & Legros 2014 (Wiley) — grounds §Limitations/spatial.
• Des Rosiers, Thériault, Kestens & Villeneuve 2002, JRER — Quebec City landscaping hedonics; now compared to our Land & Nature premium.
• Piazzesi, Schneider & Stroebel 2020, AER — segmented housing search (future work).

Unstructured data beyond text:

• Glaeser, Kincaid & Naik 2018 (NBER w25174); Poursaeed, Matera & Belongie 2018, MVA — vision-based quality extraction as sibling literature.

Interpretable ML:

• Rudin 2019, Nature MI; Lundberg & Lee 2017 (SHAP, arXiv — NeurIPS version not indexed in OpenAlex, honest @misc used); Ribeiro et al. 2016 (KDD, LIME).

NLP grounding:

• Muennighoff et al. 2023 (MTEB, EACL) — replaces the uncited "0.82 STS-B" claim.
• Yin, Hay & Roth 2019 (EMNLP) — zero-shot anchoring analogy for the reference design.

Plus replacements listed in §1 (Goodwin ×2, Zhu, Baur, Law, Alfano, Hong, Delmelle).

# 3. Sections expanded and how

• Introduction (+~1 p): stakes paragraph (mass appraisal/AVM scale, text-as-data momentum); unobserved-characteristics framing via Bajari–Benkard; "semantic dictionary" characterization; Quebec hedonic tradition; explicit three-audiences positioning; roadmap kept.
• Literature review (restructured, +~2.5 pp): new overview paragraph; §2.1 extended with origins, identification, quantile hedonics, spatial hedonics, and Quebec strands; new §2.2 "Text as Data in Economics" (dictionary→topic→embedding arc, our method placed in the Gentzkow–Kelly–Taddy taxonomy); §2.3 rebuilt around verified work in two strands (does language matter / how much information) + vision parallel; §2.4 adds MTEB and zero-shot anchoring; §2.5 reframed as a two-sided gap; comparison table rows updated to verified references.
• Methodology (+~0.3 p): dictionary lineage + zero-shot anchor paragraph in the reference-design rationale; identification paragraph now anchored to Ekeland et al./Kuminoff et al.
• Discussion (+~1.5 pp): new subsection "Relation to Prior Findings" (agreements: Nowak–Smith, Shen–Ross, Haag, Goodwin, Zietz, McMillen; departure: prediction-first literature; parallel: vision studies). Magnitude comparison added for Motivated Seller vs. Levitt–Syverson's 3.7%; Land & Nature vs. Des Rosiers et al.; Family-Friendly reframed via Kuminoff sorting + Delmelle & Nilsson.
• Conclusion (+~0.2 p): text-as-data tie-back; future-work items extended (buyer-side search via Piazzesi et al., images via Glaeser et al.).
• Data: "near-universe" claim softened to "large majority of brokered listings" (no citable source for the stronger claim).

# 4. Claims flagged for your verification

1. Levitt & Syverson 3.7% figure — resolved: verified against the published REStat paper (90(4), 599–611; agents' own homes sell 3.7% higher and stay 9.5 days longer on the market). The existing bib entry is correct.
2. Des Rosiers et al. (2002) magnitude — resolved: the paper reports landscaping premiums of roughly 4% (hedges ~3.6–3.9%, landscaped curbs ~4.4%) up to ~12.4% (landscaped patio). The discussion sentence now states this verified range ("roughly 4% to about 12%").
3. "Descriptions truncated at ~700 characters" — resolved: confirmed in the raw database. House/land/plex remarks pile up at 701–703 characters and are cut mid-word before an appended GST/QST boilerplate notice (e.g. "…z bien. GST/QST must be added to the asking price"), which is the signature of a hard excerpt cap at the source. The paper's wording ("truncated at approximately 700 characters in the data export") is accurate.
4. The softened Conneau et al. sentence ("material performance losses") replaced the previous unverifiable "up to 15%" figure.
5. The MiniLM "Spearman 0.82 on STS-B / 5× faster" claims were replaced by a softer, cited statement (Wang et al. 2020; Muennighoff et al. 2023).

# 4bis. Completion of the audit — 100% of the bibliography now verified

A final sweep verified the remaining 38 entries (the seminal econometrics and NLP works). 35 were correct as written. Three needed fixes, now applied:

• halvorsen1981choice — the entry contained the metadata of a different real paper (Halvorsen & Palmquist 1980, AER, on dummy variables in semilog equations). The paper the text relies on (choice of functional form for hedonic equations) is Halvorsen & Pollakowski 1981, Journal of Urban Economics 10(1), 37–49 — corrected.
• irwin2002interacting → irwin2002effects — the entry pointed to Irwin & Bockstael's spatial-externalities paper, but the text cites it for the value of open space; replaced with Irwin (2002), "The Effects of Open Space on Residential Property Values", Land Economics 78(4), 465–480.
• goodman1998housing → goodman1995age — the entry (Goodman & Thibodeau 1998, market segmentation) was a mismatch for the heteroskedasticity claim it supported; replaced with Goodman & Thibodeau (1995), "Age-Related Heteroskedasticity in Hedonic House Price Equations", J. Housing Research 6(1), 25–42.

With these fixes, every one of the 86 bibliography entries has been verified against OpenAlex and/or the publisher's record, and every cited claim was checked against what the cited work actually shows.

# 5. Literature potentially in tension with the paper

• Shen & Ross (2021) and Baur et al. (2023) obtain larger predictive gains from unrestricted text representations than our 20 projections — consistent with our own PCA benchmark; the paper now states this openly ("departure from the prediction-first literature") instead of claiming fit superiority.
• Delmelle & Nilsson (2021): listing text strongly encodes neighborhood composition — this supports but also sharpens the omitted-location caveat: several of our "quality" dimensions plausibly contain locational signal. The identification section already carries this caveat; it is now backed by direct evidence.
• Kuminoff et al. (2013): under equilibrium sorting, even correctly measured implicit prices are not welfare parameters — the paper's inference-scope paragraph now cites this explicitly.

# 6. Untouched

All numbers in tables/ (machine-generated from the pipeline), all figures, the data section statistics, the results and robustness sections' numeric content, and the appendices.