Source of truth
Every recall in our database has a canonical source: the official agency notice on openFDA, SaferProducts.gov, fsis.usda.gov, or api.nhtsa.gov. The full raw source data — every field the agency published, in its original JSON form — is preserved on every recall detail page. Our editorial fields (title, plain-English summary, severity score) are derived from the source. If our rewrite ever disagrees with the source, the source wins.
We do not paraphrase claims that don’t appear in the source text. We do not infer the manufacturer’s motivation. We do not speculate about who is at fault. The rewrite restates what the agency notice says, in plainer language, with the regulatory references decoded into ordinary English.
How rewrites are produced
Each rewrite is generated by a versioned prompt running against Anthropic’s Claude Haiku 4.5 model. The prompt is stored in our database (not in code), so its history is reviewable. The prompt’s instructions explicitly forbid the model from inferring facts not present in the source. Every rewrite is validated against a strict schema before being published — if the severity label doesn’t match the severity score per our rubric, the output is rejected and the record is flagged for manual review rather than published.
Prompt versions are first-class: every rewrite records the prompt version it was produced under. When we improve the prompt, the new version is applied to new ingests; existing records keep their previous version unless we explicitly reprocess them.
Severity scoring
Severity is scored 1–5 on a deterministic rubric, not by editorial judgment. The rubric uses the agency’s own classification (FDA Class I/II/III, USDA Class I/II/III) plus the presence of reported deaths or hospitalizations, the distribution scope, and the hazard type. The full rubric is published on the methodology page; the reasoning is shown in the “why this score” block on every recall page.
Manual edits
Admins can override any editorial field on any recall — for corrections, for cleanup of model artifacts, or to add a clarifying detail. Once a record is manually edited, it’s never overwritten by future LLM runs even when the prompt is bumped to a new version. The raw source data is always preserved separately and is never touched by an edit.
We don’t accept paid edits, and we don’t take edit requests from manufacturers, PR firms, or other interested parties. Corrections from manufacturers identifying factual errors are welcome but treated as ordinary correction requests evaluated against the source notice.
Corrections policy
If you spot a factual error on the site — a misspelled product name, an incorrect hazard tag, a mismatched severity score, anything — let us know. Email or web-form contact details are on the about page. We aim to acknowledge correction requests within 48 hours and publish corrections within seven days. If the correction materially changes how a reader would interpret the recall (for example, a corrected severity score), we note the correction inline.
Conflicts of interest
The Recall Desk runs no advertising, takes no affiliate commissions, and accepts no sponsored content. We have no financial relationship with any manufacturer or agency. We don’t accept gifts or free product. We don’t take payment for premium listings, removed listings, or expedited corrections.
The site is funded entirely by its operator. If that changes — we accept a grant, we add a paid feature, anything — we’ll disclose it on this page before it takes effect.
Scope and limits
We cover US federal-agency recalls only: FDA, USDA FSIS, CPSC, NHTSA, EPA. We don’t cover state-level health-department actions, international recalls, manufacturer-only safety alerts that don’t reach the federal recall channel, or product categories outside those agencies’ jurisdictions (firearms recalls, boat recalls, etc.). The sources page shows the live ingestion status for each covered agency.