# Reference verification — F4-conf-04

# Question

The draft manuscript cites five arXiv studies (R1–R5). For each one I checked the bibliographic entry and the attributed finding ("cited_for") against the extract supplied under data/sources/, which is the only admissible evidence. The three checks are: (i) does the title match? (ii) is the year consistent with the first submission or with the reproduced version? (iii) is every element of the attributed finding stated in, or arithmetically implied by, the extract? A reference that fails any of the three checks is unverifiable.

# Data and exclusions

• Inputs: data/reference_list.json (5 items) and five source extracts. Line counts are in n_input_rows. All five extract SHA-256 values match data/README.md (see out/state/data_audit.md).
• No reference was excluded (out/exclusions.csv: 5 → 5). All 5 were assessed (n_analysed = 5).
• Problems in the inputs are listed under data_issues:
  • In arxiv-2609.27732v1 (R1), the 5 km CS-DR ATT has an SE of 0.210 in Table 6 but 0.218 in Table A19.
  • The same source gives the number of control regions as 640 in one place and 608 in another.
  • The same source puts the fall in fatal outcomes at 35% in the abstract and at ~30% in Section 6.1.
  • In arxiv-2505.04555v3 (R4), the printed SE of the own-wage elasticity looks implausible.
  • The header of Table 3 in arxiv-2603.03260v2 (R3) is garbled.

# Method

out/analysis.py does the following:

1. Parses the Title header, the body title line and the "Bibliographic record" line of each extract.
2. Title check: normalises the titles (case, punctuation, diacritics, "&"/"and") and compares the entry title with both the header title and the body title.
3. Year check: takes the submission years of version 1 and of the reproduced version from the Bibliographic record. The entry year must equal one of them.
4. Finding check: splits each "cited_for" statement into its elements (magnitude, SE/CI, significance, timing, population, design). It checks each element against the specific table cell or sentence and prints that cell or sentence. Where the citation relies on a derived number, it recomputes it, for example exp(ATT)−1 or coefficient divided by the mean of the dependent variable. A finding counts as supported only if every element passes. The element-level record is written to out/finding_elements.json.
5. Baseline: a header-only screen (title and year checks only) flags baseline_header_only_unverifiable = ["R3"].

No randomness is involved (random_seed = null).

# Results

• unverifiable_references = ["R3"]; n_verified = 4. feasibility = "complete".
• R1 (Armendariz 2026): verified. Table 6, 5 km column: ATT −0.431**, SE (0.210), % change −35.0%. The "within 5 km", staggered DiD and Swiss hospital-record elements are in the abstract (reference_checks.R1).
• R2 (Li, Pearson-Merkowitz & Yokum 2026): verified.
  • The abstract states a preregistered RCT with N = 23,549.
  • Table 2, ATE column: wage $1,153 with 95% CI [166, 2,141]; reemployment 1.47 pp; weeks on UI −1.99 (reference_checks.R2).
• R3 (Liu & Zhao): unverifiable on all three checks (reference_checks.R3).
  • Title: the entry title "Restaurants, Cafés and Neighbourhood Change: Commercial Entry and House Prices in Greater London" differs from the document title "Does Entry of Food-and-Drink Establishments Raise Local House Prices? Event-Study Evidence from London".
  • Year: the entry year 2024 is neither the v1 year nor the v2 year (both 2026).
  • Finding: the 3.4–3.7% magnitude is printed, but it refers to years four and five after entry. The stacked coefficient in the event year is 0.005 (Section 3.2), and the coefficients rise afterwards (0.012, 0.018, 0.037, 0.034). The source says "prices do not jump immediately". The claim of a jump in the entry year with no further appreciation therefore contradicts the source.
• R4 (Kanayama, Miyaji & Otani 2025): verified.
  • Year: 2025 is the v1 submission year.
  • Appendix A.1, Table 2: MW elasticity −0.388 (0.132) and own-wage elasticity −0.732 (0.091). Japan's spot labour market is stated in the abstract (reference_checks.R4).
• R5 (Tossou 2025): verified. Table 3, column (4), which Section 4.1 calls the preferred specification: Treatment −0.0962 (0.0129) and % Change −0.3685, with "36.85 percent" in the text (reference_checks.R5).

# Uncertainty

This is a deterministic document check with no sampling uncertainty. Two judgement calls could affect the classification:

• R1: the SE the manuscript cites (0.210) appears in Table 6. Table A19 prints 0.218 for what appears to be the same estimate. I count R1 as supported because the cited value is printed in the main results table. A stricter reader might call the SE "internally inconsistent in the source" (data_issues).
• R4: the cited own-wage SE (0.091) matches the printed value. However, it equals the affected-employment SE, and the cross-check is inconsistent: 0.091/0.366 = 0.249 (data_issues). This points to a possible error in the source, not in the citation, so R4 is still counted as supported.

Neither call changes the R3 result, which fails on three independent grounds.

# Limitations and open issues

• The evidence is limited to the extracts. Sections omitted from them, such as Table 2 of arxiv-2510.03658 and Figures 6, 7 and 2, could not be consulted. None of the verified findings depends on them.
• R5's recomputed ratio −0.0962/0.261 = −0.3686 differs in the fourth decimal from the printed −0.3685. I attribute this to rounding of the printed inputs and do not treat it as a discrepancy.
• The source-internal inconsistencies in R1 and R4 are reported, not resolved. They cannot be resolved from the extracts alone.