Changelog

Every adverse-event report we index now carries the manufacturer's coded problem terms, mapped to the IMDRF adverse-event terminology. Previously that coding was dropped on the floor — the ingest read a field openFDA doesn't publish, so the column sat empty on all 4.6M reports.

• 8.6M problem tags backfilled across 3.68M reports — 79% of indexed events, on both the device-problem and patient-effect axes.
• New Problem Trends view: the most-reported device problems across your monitored product codes, with movement against the previous window.
• Each problem opens a detail view — manufacturer and product-code breakdown, week-by-week volume against the prior period, event-type split, and links out to the source MAUDE records.
• Signal detail and the inbox now show the coded problems behind an MDR spike, so a spike reads as what went wrong, not just how often.

Counts are reported events, not failure rates — MAUDE has no denominator. Coding is manufacturer-reported and descriptive: we surface the observational annexes only, and assert nothing about cause.

(Thanks to A.S.)

All four signal detectors now guarantee finite scores. Previously, two detector paths could emit non-numeric scores under sparse data — and because of how the database orders them, those broken signals could outrank real ones.

• Degraded inputs are excluded from ranking and logged, never silently promoted.
• Top-3 selection re-verified across all active anchors.
• Scoring methodology docs updated to match.

Replacing a company anchor is now a single atomic operation: the old firm group is removed and the new one created together, so a mid-switch failure can't leave your workspace pointed at stale monitoring data.

• Errors during anchor removal now surface in the UI instead of failing silently.
• Existing workspaces with stale groups have been repaired.

Signal detection now runs across the top 100 product codes by MDR volume, plus targeted additions requested by design partners. Adverse event history for the new codes is fully backfilled.

Firm names now roll up into corporate groups using strict prefix matching rather than fuzzy similarity. Similarity scoring occasionally merged unrelated manufacturers; prefix matching keeps groupings conservative and auditable.

The signal inbox and signal detail pages have been rebuilt on a shared design system — consistent tokens, denser information layout, and clearer severity treatment. Remaining surfaces migrate over the coming releases.

Device records are now embedded for semantic search, improving how company anchors identify the product codes relevant to a manufacturer — including catalog names that don't share obvious keywords with the FDA classification.