Argus HQ publishes four streams of FDA enforcement and regulatory data. Every public page on this site traces back to one of the sources below — we do not publish unsourced claims, and every finding on a record page is attributed to the FDA document it came from. This page documents exactly where the data comes from, how often it refreshes, what our AI enrichment does and does not do, and where the coverage is currently incomplete.
Data sources, by stream
FDA Warning Letters
Sourced from the FDA’s public Warning Letters listing at fda.gov. We scrape this listing daily at 04:00 UTC, diff it against what we’ve already ingested, and pull down any newly posted letters in full.
Drug recalls
Sourced from the openFDA drug enforcement endpoint, api.fda.gov/drug/enforcement.json. Our cron job queries this endpoint on a 10-day lookback window on every run, so a recall that posts late or gets amended by FDA is still picked up on the next cycle rather than missed permanently.
Drug approvals
Sourced from two openFDA endpoints: api.fda.gov/drug/drugsfda.json for approval records and api.fda.gov/drug/label.json for the associated label data. This is currently our newest and least complete stream — see Known limitations below.
Form 483 inspection observations
FDA does not currently offer a comprehensive public API for Form 483s. We source what we can from FDA’s public inspection databases, and we are disclosing openly that 483 coverage is pending an API key application with FDA and is materially incomplete today. Any 483 record we do publish is reproduced verbatim from a document FDA has already made public — we do not estimate, infer, or synthesize 483 content.
What our AI enrichment is, and isn’t
Every record page reproduces its FDA source document in full, verbatim. Above that, we add a short analysis section. That analysis section is AI-drafted directly from the source record and reviewed against an automated fact-check gate before it publishes. The gate checks the draft against the source document it was generated from and blocks publication if it introduces a claim, company name, or citation not present in the source. This means: the AI is a summarization and formatting layer, not an independent research process. It does not add outside facts, does not speculate about outcomes or intent, and does not offer legal, financial, or medical advice. If you ever find an analysis section that contradicts or misstates the underlying FDA document, that’s a bug — see Corrections below.
Corrections
If you find an error — a misread date, a misattributed company, a broken source link, an AI-drafted summary that misstates the source document — email corrections@argushq.ai with the page URL and what’s wrong. We fix confirmed errors within 2 business days and add a dated correction note directly on the affected page. See the Editorial Policy for the full corrections process.
Known limitations
- Form 483 coverage is pending. We do not yet have an FDA API key for inspection-citation data, and 483 records on this site are limited to what is separately public. This is the single largest coverage gap in the product today.
- Drug approval coverage is partial. The approvals stream is the newest of the four and does not yet cover every NDA/ANDA/BLA action FDA has taken; we are expanding it incrementally.
- Recalls reflect a 10-day lookback per run. A recall that FDA amends more than 10 days after its initial posting may not be re-captured until we widen that window.
- Entity matching is automated and imperfect. We match companies named in enforcement records to entity profiles algorithmically. Misattribution is possible, particularly for companies with similar names or multiple corporate subsidiaries.
See Data Sources for a table view of refresh frequency and fields captured per stream, and the Updates log for a running, dated record of what was added and when.
Dataset paper
For a citable, arxiv-style writeup of the dataset’s methodology, structure, and statistics, see the Argus HQ FDA Enforcement Dataset paper (PDF). Bulk licensing for enterprise use is described on the data licensing page.
Programmatic access
Every stream described above is also available programmatically -- an MCP server for Claude/ChatGPT/Cursor, four per-entity JSON APIs, an OpenAPI 3.1 spec, and an llms.txt file. See /developers for install instructions and examples.