HOW A SIGNAL BECOMES AN ALERT
A source record becomes an alert only after it clears that watch's similarity gate. The source facts, rationale, severity, and your decision stay separate.
Trademark alert
HIGHIllustrative record · not a real filing
A live watch checks its named public source on its published schedule.
Source-specific signals decide what is close enough for human review.
The source, rationale, severity, and your decision remain separate.
A new record that clears a similarity gate.
A high- or moderate-severity candidate sent for review.
The disposition you record without rewriting the evidence.
Normalized text, phonetic, token, and class signals narrow the daily USPTO application feed.
Class and title narrow new grants before visual triage compares the drawing with your registered image.
Character confusables, edit distance, and phonetic matching evaluate newly delegated names.
What we do read, and how often, is on the coverage page.
Candidate → alert → your call.
A candidate is a new filing (or, for a domain watch, a newly observed name) that clears the similarity gate for one of your watched assets. An alert is a candidate whose source-specific decision produces a high or moderate verdict. Trademark and design watches use AI triage; domain watches use deterministic signal thresholds. The alert is emailed to you and stored on your dashboard with the evidence it was built from. What you do next is your call: MarkSentry monitors and evaluates; it does not file, oppose, or take anything down on your behalf. On your dashboard you record your call on each alert (Relevant or Not a concern), and the review rail empties as you do.
Before any AI is involved, a deterministic gate decides whether a new trademark filing is worth comparing at all. Both marks are normalized first (lower-cased, split on anything that is not a letter or digit, and joined back into one canonical string), and then exactly these signals are computed between them:
A filing that passes any one of the three similarity signals (and the class filter) is a candidate. Candidates are ordered by edit distance and the closest 25 go to triage in one call, never more, so a noisy day cannot run up an unbounded bill.
Design watches gate differently: a class filter on the grant's USPC and Locarno codes, then the same three text signals (edit distance, phonetic match, token overlap) on your product label against the grant's title. Text only, which catches two mitten patents, not two mittens that look alike. The closest ten by title edit distance then go to vision triage, which is the real signal for designs.
Domain watches do not call an AI triage model. Their deterministic gate treats an exact or confusable-folded brand-token embed as high severity, a Metaphone match within edit distance two as moderate, and everything else as none. The stored rationale is built from those same computed signals.
The trademark triage prompt asks, in its own words, for a judgment on appearance, sound, meaning/commercial impression, and goods/services class overlap, and defines the four severity bands exactly like this:
The design-patent vision triage uses the same four band names, defined for drawings; its top band reads "high" (an ordinary observer could mistake one design for the other).
High and moderate become alerts. Low is shown on a free-scan report but never emailed; none is dropped from both.
The coverage line on a report or dashboard is a measured ingest watermark, never a configured promise. Trademark coverage uses the earliest and latest daily USPTO files actually read plus the distinct-day count. Design coverage is measured separately by weekly issue. Domain coverage is measured separately for each approved TLD from its zone-file watermark; certificate-transparency signals and skips are recorded separately and are not counted as zone-file coverage. Missing records are stated rather than filled in.
Monitor + Guided.
Every live watch is monitor-and-guided: we read the source, apply that watch's source-specific decision, and show you the record, rationale, verdict, and timeline context when one exists; you decide. Trademark and design candidates use AI triage; domain candidates use deterministic severity. No MarkSentry watch submits anything to a registry, marketplace, or registrar on your behalf.