# Reference verification report — F4-conf-04

# Question

For each of five references (R1–R5) in a draft manuscript, does the provided source extract confirm (i) the title, (ii) the year (first-submission year or year of the reproduced version) and (iii) the attributed finding ("cited_for")? A reference that fails any check is unverifiable.

# Data and exclusions

Inputs: data/reference_list.json (5 items) and five arXiv extracts; line counts are in n_input_rows (README.md 17, arxiv-2609.27732v1 134, arxiv-2606.14621v1 58, arxiv-2603.03260v2 107, arxiv-2505.04555v3 66, arxiv-2510.03658v1 91 lines). All SHA-256 values match data/README.md (see out/state/data_audit.md). No exclusions were applied (out/exclusions.csv: 5 in, 5 analysed; n_analysed = 5). The only evidence used is the extracts, as the task requires.

# Method

out/analysis.py reads each extract's header (Title, Bibliographic record) and checks that the body title line matches it. It then splits each entry into title, year and arXiv number.

• Title: compared after normalisation (case, punctuation, diacritics, "&"/"and").
• Year: must equal one of the submission years in the Bibliographic record.
• Finding: each element of cited_for is checked against the specific table cell or sentence. The script prints these, along with arithmetic checks where relevant.

Baseline (P2): a check of title and year only (baseline_unverifiable_title_year_only, baseline_n_verified_title_year_only). No randomness is involved (random_seed = null).

# Results

• unverifiable_references = ["R3"]; n_verified = 4.
• R1 (reference_checks.R1): passes all three checks. Table 6, 5 km column: ATT -0.431** (0.210), -35.0%. The abstract describes a staggered DiD on Swiss hospital records.
• R2 (reference_checks.R2): passes all three. Table 2, ATE column: $1,153 (95% CI [166, 2,141]), 1.47 pp, -1.99 weeks. The abstract gives a preregistered RCT with N = 23,549.
• R3 (reference_checks.R3): fails all three checks.
  • 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 says 2024; the record shows submissions in 2026 only (v1 3 Mar 2026, v2 2 Apr 2026).
  • Finding: the stacked coefficient is 0.005 at the event year and rises to 0.037/0.034 at τ=4/5 (Section 3.2; Table 3, Clean-onset row). The source therefore describes a gradual rise, not a 3.4–3.7% jump in the entry year with no further appreciation.
• R4 (reference_checks.R4): passes all three. The year 2025 equals the v1 year. Appendix A.1 Table 2: -0.388 (0.132) and -0.732 (0.091).
• R5 (reference_checks.R5): passes all three. Table 3 column (4), the preferred specification: -0.0962 (0.0129), % Change -0.3685, which matches the 36.85% stated in Section 4.1.
• Baseline comparison: the title/year-only check also flags only R3 (baseline_unverifiable_title_year_only). The finding check adds no further flags.

# Uncertainty

The checks are deterministic, so there is no sampling uncertainty. The uncertainty that remains is about judgement:

• R1: the source gives two different standard errors for the same specification, (0.210) in Table 6 and (0.218) in Table A19 (data_issues). The cited 0.210 matches the main-results table, and 5% significance holds under both. I treated it as supported.
• R4: the 0.091 SE of the own-wage elasticity is identical to the SE of "Affected employment". It may be a duplication in the source, but the cited values are printed verbatim.
• R5: -0.0962/0.261 = -0.3686, while the table prints -0.3685. This is a rounding difference.

None of these changes the set of unverifiable references.

# Limitations and open issues

• The extracts are partial. Findings are judged only against what the extracts contain.
• Source inconsistencies are listed in data_issues and left unresolved:
  • R1 source: SE 0.210 vs 0.218, and 608 vs 640 control regions; the abstract says fatal outcomes fall 35%, Section 6.1 says about 30%.
  • R4 source: the duplicated SE; the note says "about 4%" while the table gives 0.047.
  • R5 source: the abstract's 0.0855/18.79% refers to a different table.
• Under a stricter reading, the R1 SE discrepancy or the R4 duplicated SE could be treated as grounds to flag the reference. I did not, because the cited values appear verbatim in the extracts.
• feasibility = "complete".