# F4-conf-04 — Verification of five references against source extracts

## Question
For each reference R1–R5 in a draft manuscript, does (i) the title match the document, (ii) the year agree with the first-submission year or the reproduced-version year, and (iii) the attributed finding ("cited_for") appear in, or follow arithmetically from, the provided arXiv extract? References failing any check are unverifiable.

## Data and exclusions
Inputs: `data/reference_list.json` (5 items) and five source extracts (line counts in `n_input_rows`); `data/README.md` was read for provenance. All SHA-256 values match README and bundle (see `out/state/data_audit.md`). No reference was excluded (`out/exclusions.csv`); all 5 were assessed (`n_analysed` = 5). Only the extracts were used as evidence; years come only from each extract's "Bibliographic record" line.

## Method
`out/analysis.py` (deterministic, no randomness, `random_seed` = null) parses each entry. It compares the normalised title (case, punctuation, diacritics and "&" ignored) with both the header and the body title line. It checks the entry year against the first-submission year and the reproduced-version year. It then ties each element of the cited finding to specific printed table cells or text lines, and prints those cells. Arithmetic implications are recomputed (e.g. exp(−0.431)−1 = −35.0%; −0.268/0.366 = −0.732; −0.0962/0.261 = −0.3686). Baseline: a check of title and year only. It flags the same set as the full check (`baseline_metadata_only_unverifiable` = ["R3"]).

## Results
Per-reference record: `reference_checks`.
- **R1** (Armendariz 2026): all three checks pass. Table 6, 5 km column, prints ATT −0.431** (0.210), −35.0%. The text states significance at 5%, and the abstract describes the staggered DiD on Swiss hospital records (`reference_checks.R1`).
- **R2** (Li, Pearson-Merkowitz & Yokum 2026): all checks pass. Table 2 ATE column gives $1,153 [166, 2,141], 1.47 pp and −1.99 weeks; the abstract gives the preregistered RCT and N = 23,549 (`reference_checks.R2`).
- **R3** (Liu & Zhao): **fails all three checks** (`reference_checks.R3`).
  - Title: the entry's title ("Restaurants, Cafés and Neighbourhood Change: …") is not the document title ("Does Entry of Food-and-Drink Establishments Raise Local House Prices? Event-Study Evidence from London").
  - Year: 2024 is neither the v1 year nor the v2 year (both 2026).
  - Finding: the extract reports 0.005 at τ=0, rising gradually to 0.037 (τ=4) and 0.034 (τ=5), and states that "prices do not jump immediately". This contradicts the claim of a 3.4–3.7% jump in the entry year with no later appreciation.
- **R4** (Kanayama, Miyaji & Otani 2025): all checks pass. 2025 is the v1 year. Appendix Table 2 prints −0.388 (0.132) and −0.732 (0.091) for Japan's spot labor market (`reference_checks.R4`).
- **R5** (Tossou 2025): all checks pass. Table 3 column (4), which the text calls the preferred specification, prints −0.0962 (0.0129) and % Change −0.3685, described in the text as 36.85% (`reference_checks.R5`).

Unverifiable set: `unverifiable_references` = ["R3"]; `n_verified` = 4; `feasibility` = "complete".

## Uncertainty
This is a deterministic document check with no sampling uncertainty. The judgement calls are:
- **R1:** the SE differs across tables (0.210 in Table 6 vs 0.218 in Table A19). The citation matches the main table, and 5% significance holds under either value.
- **R5:** the abstract headline (0.0855, 18.79%) refers to a different outcome (Table 2, omitted from the extract) than the cited Table 3 binary outcome.
- **R3:** the verdict does not depend on any judgement call, since all three checks fail independently.

## Limitations and open issues
- Source-internal inconsistencies are listed in `data_issues`:
  - R1: SE 0.210 vs 0.218, and 608 vs 640 control regions; the abstract's 35% / "comparable" vs Section 6.1's 30% / 47%.
  - R4: the own-wage elasticity SE is identical to the affected-employment SE.
  - R5: the abstract and Table 3 differ, and footnote 11 is garbled.
  - Rendering defects: R2 and R3.

  These issues cannot be resolved from the extracts. None of them makes a cited element absent from the extract, so they are recorded rather than counted against the references.
- Only the excerpted sections were available. A finding located only in omitted sections would have been judged unsupported, but no such case arose.