bibcheck — Per-Entry Bibliography Audit

Audit-only sister skill to /cite. Given a .bib file, spawn one subagent per entry to verify the entry against canonical sources (DOI metadata, Semantic Scholar, OpenAlex, arXiv, Zotero). Report PASS / WARN / FAIL per entry with a one-line diagnostic, group by severity, and write a dated report plus a corrected.bib the user can review and merge themselves. Never overwrites the source .bib.

Why per-entry parallelism. One agent asked to audit 80 entries in a single pass drifts: early entries get careful treatment, late entries get pattern-matched. Spawning one subagent per entry gives each citation a full attention budget and removes the late-batch quality cliff. This mirrors the structure of /seven-pass-review (parallel lenses) and /audit-reproducibility (per-claim checking).

The LLM-fabrication concern. Hallucinated citations have a recognizable signature: plausible author names + plausible title + a DOI that 404s + no Semantic Scholar / OpenAlex / arXiv match anywhere. /bibcheck is built to surface exactly this pattern.

When to use

• Pre-submission: before MKSCI / JMR / JCR / MS submission or resubmission, especially for drafts where any text was AI-assisted.
• Inherited `.bib`: when picking up a coauthor's or RA's bibliography and you don't know its provenance.
• Post-Zotero-export sanity check: even Better BibTeX can carry forward bad upstream metadata.
• After a large `/litreview` batch: verify that everything the lit-review skill staged into .bib actually exists.

Inputs

• $0 — path to a .bib. If omitted, Glob the cwd for references.bib, refs.bib, bibliography.bib, then any *.bib. Overleaf projects typically live under <OVERLEAF_ROOT>/<PROJECT_SUBDIR>/. If multiple .bib files exist, grep the main .tex for \bibliography{} / \addbibresource{} to pick the right one; ask the user if still ambiguous.

Workflow

Phase 0: Parse the .bib

1. Read the full .bib.
2. Split into entries on ^@\w+\{ boundaries. Capture entry type (article, book, incollection, misc, inproceedings, unpublished), citation key, and every field = {value} pair (handle braced and quoted forms, escaped braces, multi-line values).
3. Build an in-memory list of entries. Note malformed entries (unbalanced braces, missing key, duplicate key) — these go straight to FAIL without dispatching a subagent.
4. Create a working directory <bib_dir>/bibcheck_<YYYYMMDD_HHMM>/ with subfolders entries/ (one file per entry for subagent input) and reports/ (one JSON per entry for subagent output).

Phase 1: Spawn per-entry subagents

For each well-formed entry, launch a Task subagent with a focused prompt: "Verify this single BibTeX entry against canonical sources. Return JSON with verdict and per-field findings." Cap concurrent subagents at 8 (--max-parallel if the user passes it). Each subagent runs the per-entry checks below and writes one JSON file to reports/<citekey>.json. Lower the cap if rate-limited; raise it for very large bibliographies.

Phase 2: Per-entry checks (run by each subagent)

In order, stopping early once verdict is determined:

1. DOI resolution. If the entry has a doi field:

• Call mcp__semantic-scholar__get_paper_details(paper_id="DOI:<doi>"). On hit, compare title / author surnames / year / venue / volume / issue / pages against the entry. Discrepancies → field-level WARN; outright mismatches → FAIL.
• If S2 misses, fall back to WebFetch("https://api.crossref.org/works/<doi>") for Crossref ground truth.

1. arXiv resolution. If the entry looks like a preprint (eprint = {2403.12345}, archivePrefix = {arXiv}, or note = {arXiv:...}):

• Call mcp__arxiv__get_abstract(paper_id="<id>") to confirm existence and pull canonical metadata.
• Cross-check whether the same paper now has a published version with a DOI (use mcp__semantic-scholar__search_papers(query="<title>")) — if so, emit a WARN suggesting the published @article entry instead of the preprint @misc.

1. Title-only search (no DOI, no arXiv). Call mcp__semantic-scholar__search_papers(query="<title>", limit=5). Fuzzy-match against the entry's title (case-insensitive, normalize punctuation, accept ≥ 0.85 cosine on token-set). On a confident hit, compare authors / year / venue. On no plausible match, also try mcp__openalex__openalex_search_entities(entity_type="works", query="<title>") as a second opinion.
2. Fabrication test. If (a) doi is present but DOI doesn't resolve in S2 OR Crossref, AND (b) title search returns no fuzzy match in S2 OR OpenAlex OR arXiv → FAIL with diagnostic fabricated? — DOI does not resolve and no metadata source recognises the title. This is the LLM-hallucination signal.
3. Author normalization. Compare author surnames as the ground-truth source returns them. Flag:

• Wrong surname (clear mismatch) → FAIL.
• Initials vs. full first names → no flag (style choice).
• Missing accents (Pena vs. Peña), missing tussenvoegsels (Van Den Bulte vs. van den Bulte) → WARN.
• Author-list truncation (& others or missing trailing authors) → WARN.

1. Year reconciliation. If the entry year differs from the canonical publication year:

• If eprint/archivePrefix=arXiv present and entry year matches arXiv-submission year → WARN ("entry cites preprint year <y1>; published version is <y2>"). Do not auto-FAIL — citing the preprint year is a legitimate choice.
• Otherwise → FAIL ("year mismatch: entry says <y1>, canonical says <y2>").

1. Journal canonicalization. Compare journal against canonical name from S2 / OpenAlex. Marketing-relevant cases worth WARNing on:

• Abbreviation drift: J. Mark. Res. vs. Journal of Marketing Research, Mark. Sci. vs. Marketing Science, Mgmt. Sci. vs. Management Science, Quant. Mark. Econ. vs. Quantitative Marketing and Economics.
• Wrong journal entirely (different ISSN family) → FAIL.
• Apply the same logic for booktitle on @inproceedings (ICML, NeurIPS, KDD — full names).

1. Volume / issue / pages. Compare numeric fields. WARN on a single-field mismatch; FAIL if 2+ disagree (likely a swapped-entry bug — title of paper A glued to the volume/pages of paper B).
2. Zotero cross-check (optional). If a citation key looks Better-BibTeX-shaped (e.g., lastname2024Word), call mcp__zotero__zotero_search_items(query="<title>", qmode="titleCreatorYear", limit=3). If a Zotero hit exists with a different DOI than the entry, WARN. The Zotero MCP is read-only — never attempt to write back; surfacing the mismatch is sufficient.

Each subagent emits:

{
  "citekey": "smith2024paper",
  "verdict": "PASS | WARN | FAIL",
  "diagnostic": "one-line summary",
  "field_findings": [
    {"field": "year", "status": "FAIL", "entry_value": "2023", "canonical_value": "2024", "source": "doi.org/10.1287/mksc.2024.0123"}
  ],
  "canonical_source": "doi | s2 | openalex | arxiv | none",
  "suggested_replacement": "<full corrected BibTeX entry, or null if PASS / unfixable>"
}

Phase 3: Aggregate

Read every JSON in reports/. Bucket entries by verdict: FAIL > WARN > PASS. Within each bucket, sort by citekey alphabetically (stable, easy to diff across runs).

Phase 4: Report

Write <bib_dir>/bibcheck_<YYYYMMDD_HHMM>/bibcheck_report.md:

# Bibliography Audit: <bib filename>

**Date:** YYYY-MM-DD HH:MM
**Source:** <full path to .bib>
**Entries audited:** N (parsed) / M (malformed, skipped)

## Summary
| Verdict | Count |
|---|---|
| PASS | a |
| WARN | b |
| FAIL | c |
| Possibly fabricated | d  (subset of FAIL)|

## FAIL (blockers — verify before submission)
| Citekey | Diagnostic | Canonical source |
|---|---|---|
| smith2023paper | year mismatch: entry 2023, canonical 2024 | doi.org/10.1287/mksc.2024.0123 |
| jones2024nope  | fabricated? DOI 404s, no S2/OpenAlex/arXiv match | none |

## WARN (style / preprint-vs-published / minor field drift)
| Citekey | Diagnostic |
|---|---|

## PASS
<one-line per entry>

## Per-entry detail
<expandable section with the JSON for each FAIL and WARN>

## Next steps
1. Review each FAIL row. Open `corrected.bib` for suggested replacements.
2. Diff `corrected.bib` against the source `.bib` (`git diff --no-index`).
3. After review, move `corrected.bib` into place yourself — this skill does not overwrite the source.

Also write <bib_dir>/bibcheck_<YYYYMMDD_HHMM>/corrected.bib: a full copy of the source .bib with FAIL/WARN entries replaced by their suggested_replacement (preserving every PASS entry untouched, preserving comments and entry order). Entries with no confident correction are kept as-is and tagged with a leading % bibcheck:FAIL — <diagnostic> comment line.

Pass / Warn / Fail criteria

• PASS — DOI or canonical source resolves; title, authors, year, venue, volume/issue/pages all match within style tolerance.
• WARN — at most one of: author accent/tussenvoegsel drift, journal abbreviation drift, preprint-year vs. published-year, single volume/issue/page disagreement, preprint entry where a published version now exists.
• FAIL — any of: DOI doesn't resolve, wrong surname, wrong year (not preprint-explained), wrong journal, two-or-more volume/issue/page disagreements, or the fabrication signature (no source recognises the entry).

Output report shape

Files written to <bib_dir>/bibcheck_<YYYYMMDD_HHMM>/:

input.bib              # verbatim copy of source for provenance
entries/<key>.bib      # one parsed entry per file (subagent inputs)
reports/<key>.json     # one verdict per file (subagent outputs)
bibcheck_report.md     # consolidated human-readable report
corrected.bib          # drop-in candidate replacement (user reviews + merges)

Relation to /cite (do not duplicate)

If /bibcheck finds an entry that should be replaced and the user wants to do that via /cite, hand off the DOI / arXiv ID to /cite rather than mutating the .bib here.

Long batch runs

Bibliographies over ~60 entries with --max-parallel 8 can take several minutes. Use Bash run_in_background: true for the dispatch loop and stream subagent completion via Monitor. Do not block on a single slow MCP call — set a 30-second per-entry timeout and emit a WARN: lookup_timeout rather than hanging.

Failure modes

• `.bib` not found. Ask the user for the path; do not invent one.
• All entries return FAIL with `lookup_timeout`. MCP rate-limit or network issue. Pause, retry once with --max-parallel 2. If still failing, surface the error to the user rather than silently marking everything FAIL.
• Many false-positive FAILs on `year`. Likely a preprint-vs-published mismatch pattern; widen Phase 2 step 6's preprint detection rather than loosening tolerance.
• Match confidence universally below 0.7. The .bib may be a non-English-titled set or a niche venue not indexed by S2/OpenAlex. Flag and ask the user to point at a sample entry known to be correct, then recalibrate.

Out of scope (do not do these)

• Don't add entries. That's /cite. If an entry should be replaced, write the suggestion to corrected.bib; do not POST to Zotero.
• Don't auto-overwrite the source `.bib`. Ever. The user diffs and merges themselves.
• Don't claim a citation supports a textual claim. That's a /review-paper literature-fidelity check; /bibcheck only verifies the entry's metadata matches a real paper.
• Don't bulk-reformat. Entries that PASS are copied verbatim into corrected.bib, even if the user's style preference differs.
• Don't audit `.bib` files in `_archive/` or `Old/` subdirectories unless explicitly asked — these are intentionally frozen.

Cross-references

• /cite — adds individual entries; complementary forward direction.
• /litreview — finds papers to cite; pair before /cite and /bibcheck.
• /audit-reproducibility — same per-claim subagent-fanout structure, applied to numeric claims rather than citations.
• /review-paper / /seven-pass-review — broader manuscript reviews; the citation-pass in those skills delegates here.