Methodology
How severity is scored
Every recall on this site is scored from 1 to 5 against the rubric below. The score is determined from the agency’s own classification, the presence of documented injuries or deaths, distribution scope, and hazard type — never by editorial judgment about how worrying a recall sounds.
The rubric
| Score | Label | Criteria |
|---|---|---|
| 5 | Critical | FDA Class I; NHTSA “death/serious injury reported”; or known fatalities. |
| 4 | Severe | FDA Class II with hospitalization reports; USDA Class I; structural vehicle defect; or significant injury / property-damage reports. |
| 3 | High | Allergen mislabeling on common allergens; high-risk pathogens (Listeria, E. coli O157, Salmonella) without reported illness; risk-of-harm products with no injury reported yet. |
| 2 | Moderate | Minor labeling errors, low-risk contamination, voluntary precautionary recalls. |
| 1 | Low | Cosmetic defects, documentation issues, packaging-only problems. |
Hard rules
- If reported deaths exist anywhere in the source text, the score is 5 regardless of other factors.
- If the agency’s own classification is FDA Class I or USDA Class I, the score is at least 4.
- If the source text explicitly states no illnesses/injuries reported and the hazard is theoretical, the score is at most 3.
- FDA / USDA Class III recalls are typically 1 or 2 — they are by definition unlikely to cause adverse health consequences.
LLM rewrite process
- Our ingestor fetches the raw recall notice from the agency’s public API and stores it verbatim alongside the rewritten text. The original JSON is preserved so the rewrite is always re-derivable.
- A Claude Haiku 4.5 prompt — versioned and stored in the database, not in code — reads the raw text and the rubric above and produces JSON output for the editorial fields. The prompt instructs the model not to infer facts that aren’t in the source.
- We validate the output against a strict schema (Zod) and reject any rewrite where the severity label doesn’t match the severity score per the rubric. Validation failures are flagged for admin review rather than published.
- Admins can override any field manually. Once a record is manually edited, it’s never overwritten by future LLM runs, even when the prompt is bumped to a new version.
- Prompt versions are first-class: every rewrite records the prompt version it was produced under, so a future improvement can selectively reprocess older records.
When in doubt, click through
Every recall page on this site links back to the official agency notice. If a number, lot code, or remedy instruction is critical to your decision, read it from the source. Our rewrite exists to help you find the recall and understand it quickly — not to replace the agency’s own authoritative copy.