# F4-conf-04 — Reference-list verification

# Question

For each of five references in a draft manuscript, I checked the entry against its source extract under data/sources/ on three points: (i) does the title match, (ii) is the year consistent with the arXiv submission dates, and (iii) is the finding the manuscript attributes to it ("cited_for") stated in the extract or arithmetically implied by it? The goal is the set of references that cannot be verified.

# Data and exclusions

• Inputs: data/reference_list.json (5 items), five source extracts (66–134 lines each) and data/README.md (17 lines). The line counts are in n_input_rows.
• The only evidence used was the extracts: their headers ("Title", "Bibliographic record") and their bodies. I did not use any external bibliographic knowledge or any other versions of the papers.
• No references were excluded. All 5 were assessed (n_analysed = 5).

# Method

• Title: I parsed the entry title and compared it with both the header title and the body title. Before comparing, I normalised case, punctuation, diacritics, "&"/"and" and -isation/-ization spellings.
• Year: an entry year is consistent if it equals the year of the first submission or the year of the version reproduced in the extract. Both years come from the "Bibliographic record" line.
• Finding: I matched every element of "cited_for" (magnitude, SE/CI, significance, timing, population/design) to a printed table cell or a sentence in the text. Where an element could be derived arithmetically, I recomputed it (e.g. exp(ATT)−1, coefficient/mean, ratio elasticities).
• out/analysis.py prints every header line, table row and sentence it relies on, and regenerates out/results.json.

# Results

Unverifiable references: R3 (n_verified = 4). R3 fails all three checks:

• Title: the entry's title ("Restaurants, Cafés and Neighbourhood Change: Commercial Entry and House Prices in Greater London") is a different title from the document's ("Does Entry of Food-and-Drink Establishments Raise Local House Prices? Event-Study Evidence from London").
• Year: 2024 matches neither submission year (both 2026).
• Finding: the extract describes a gradual rise, from about 0.5% in the event year to 3.4–3.7% by years four and five, and states explicitly that "prices do not jump immediately". The manuscript's attribution (a 3.4–3.7% jump in the entry year with no further appreciation later) contradicts this.

Arithmetic cross-checks behind the supported findings:

• R1: exp(−0.431)−1 = −35.0%, and |t| = 2.05.
• R4: affected employment / affected wage = −0.268/0.366 = −0.732. Also, (missing + excess)/%ΔMW = −0.383, against the printed −0.388; I attribute this small gap to rounding of the printed inputs.
• R5: −0.0962/0.261 = −0.3686, against the printed −0.3685.

# Uncertainty

This is a deterministic, document-based check, so there is no sampling uncertainty. The judgment calls are:

• R1: the extract contradicts itself on the SE. Table A19 prints (0.218) for the same CS-DR 5 km estimate that Table 6 prints with SE (0.210), and the A19 note gives 640 control regions where Tables 6–7 give 608. The cited SE matches Table 6, the main results table, so I marked the finding supported. The 5% significance claim holds under either SE (|t| = 2.05 or 1.98). A stricter reviewer might ask the author to reconcile the two numbers.
• R4: the own-wage elasticity's SE (0.091) is identical to the SE of "Affected employment". That may be a duplication in the source, but the citation reproduces the printed values exactly.

# Limitations and open issues

• The extracts reproduce only selected sections. A finding is judged only against what the extract contains.
• Other internal inconsistencies I noticed but that do not affect any cited finding (listed in data_issues):
  • In the Armendariz abstract, fatal hospitalisations "fall by 35%", while Section 6.1 says ~30% and Table 7 implies −29.8%.
  • In Tossou, footnotes are rendered inline, and Section 4.1 refers to a Table 2 that is not in the extract.
  • In Liu & Zhao, the Table 3 header is garbled; I used the column order from the editor's note.
  • In Li et al., there is a "p ¡ 0.001" glyph defect.
• I treated author formatting and British/American spelling ("hospitalisation", "randomised") as non-substantive, as TASK.md instructs.