Cito, Tuto, iucunde: swiftly, safely, pleasantly. Cito finds the literature. Tuto makes sure it is cited correctly.

A citation-integrity audit pipeline, plus a public integrity report for major venues. First run: the ACL 2026 proceedings (4,459 papers, main conference and Findings).

We checked all 209,985 references in ACL 2026 on two levels: whether the cited work exists (L1), and, for claim citations, whether the cited paper actually supports the claim it is attached to (L2).

• Site: tuto.fim.ai, the public report; the single-paper self-service checker is at /check
• Report source: docs/REPORT-acl-2026.en.md
• Dataset: dataset/, 18,724 + 3,795 citation-level judgments and 169 arbitration records, anonymized, CC BY 4.0; schema in its README
• Docs: docs/PRD.md (product spec), docs/ARCHITECTURE.md (technical architecture)

Reference lists are not in any metadata. anthology.bib carries a paper's own metadata but not who it cites. Second-hand citation databases do not have it either: Crossref reports reference-count = 0 for ACL papers, and OpenAlex reports referenced_works = 0.

The deeper reason is that those databases only store references that matched something. A fabricated reference is, by construction, the one that failed to match and was silently dropped. Hunting hallucinated citations in a second-hand citation database is logically self-defeating. You have to read back the actual string the author wrote, from the PDF.

Single-paper checking is live at tuto.fim.ai/check. Drop in an arXiv ID and it runs the same pipeline as the full audit (existence, claim support, arbitration), returning leads for human review in a few minutes. The server lives in src/tuto/check/ (FastAPI, pip install -e ".[api]") and self-hosts on one machine: uvicorn tuto.check.service:app plus a GROBID container and a DBLP snapshot.

ingest → parse → verify(L1) → triage(automated review funnel) → arbiter(L2 arbitration) → report

Humans do only three things: spot-check a sample to compute the false-positive rate, adjudicate the suspicious list (12 items, verified one by one), and select and anonymize the cases that appear in the report.

uv venv --python 3.12 && uv pip install -e .

# Ingest: all bib entries and PDFs (rate-limited to ~1MB/s per IP;
# 4,459 papers takes about 3 hours, resumable)
uv run tuto ingest --venue acl-2026

# Parse: GROBID → refs.jsonl + contexts.jsonl, plus a parse-acceptance sample
uv run tuto parse --venue acl-2026 --grobid-url http://localhost:8070

GROBID needs to be running: docker run -d --name grobid -p 8070:8070 lfoppiano/grobid:0.8.1

To audit a different venue, add a row to the VENUES table in src/tuto/ingest/acl_anthology.py.

Every L1 existence check runs against a local DBLP snapshot, so a stale one shows up as false "could not find" leads on recent papers. dblp.org rebuilds its dump nightly and publishes an md5 beside it, so the refresh asks what changed before pulling a gigabyte:

uv run tuto refresh-dblp              # rebuild only if upstream moved
uv run tuto refresh-dblp --force      # rebuild regardless

It downloads dblp.xml.gz, verifies the published md5, parses it into a temporary database, and only then os.replace()s it onto data/cache/dblp/dblp.sqlite. The swap is atomic and a build that parses implausibly few records is refused, so the live /check service keeps serving from the old snapshot through a refresh or a failed one. Concurrent runs take a lock and exit. A full rebuild is ~8.6M records and takes tens of minutes.

In production this runs monthly from cron against the API container:

30 4 1 * * /usr/bin/docker exec tuto-api tuto refresh-dblp --quiet >> /root/tuto-data/refresh-dblp.log 2>&1

• Data source: ACL Anthology directly (full bib dump, PDFs at predictable URLs)
• Parsing: GROBID processReferences. Not MinerU. ACL PDFs are born-digital LaTeX output with no OCR need; MinerU and OCR are reserved as a scanned-document fallback for the V2 upload tool
• Parse acceptance: a layout-based reference counter working on a different principle from GROBID (it counts flush-left lines of hanging-indent blocks) cross-checks the whole corpus. Only the disagreements get manual review, which yields a recall estimate with an interval, published
• L1 verification: local snapshots first (DBLP / arXiv / OpenAlex / Anthology bib + the Cito index), falling back to the Crossref and OpenAlex APIs
• Hard rule: the public report carries aggregate statistics and anonymized cases only. It never names a paper or an author. There is no author notification and no appeal process (see report §3.4)

cd web && pnpm install && pnpm dev   # http://localhost:5297

The site reads the report Markdown from ../docs/ at build time.

• Code: Apache-2.0 (see LICENSE)
• Report text (docs/REPORT-*): CC BY 4.0