Marksweep (Vanta Point Labs)

Marksweep: a photo from the market stall in, a trademark evidence record out. Google Lens tells you what a logo is; Marksweep tells a lawyer why it might be infringing, with a score, a reason and a chain of evidence. Brand owners lose the fight against copies at the point of sale. Investigators photograph markets, shop windows and shelves; paralegals then sort hundreds of photos by hand, and nothing records which mark was compared with which, on what basis, or when. Marksweep turns one field photo into a hashed evidence record in under a minute: SAM 3.1 isolates every logo, label and pattern (boxes and lossless masks with stable object ids); Muse Spark 1.3 reads each crop next to the office drawings of the client's registered marks and returns schema-checked JSON with six sub-scores and one-sentence rationales on the likelihood-of-confusion global assessment (visual, aural, conceptual, goods, distinctiveness, imitation signals), a short narrative and the single most relevant registration. The record keeps the photo and crop hashes, boxes, scores, rationales, field notes and a ledger of every model call (model, purpose, latency, tokens), with click-through links to USPTO, EUIPO and WIPO. Every number is a research signal for the supervising lawyer; the tool never says "infringes". If no client mark is a candidate, every score is capped at 25 and the narrative says so (see the control demo). Marksweep is the field-intake module of Signum, our AI-assisted IP case assessment piloting with a Hong Kong IP firm. Live demo: https://meta.vantapointlabs.com (four recorded sweeps replay instantly; an operator token runs new photos live). Everything in the repo was built on 4 November 2026; reused and labelled: six dimension names and weights from public case law and eight public register drawings used as the demo clients' portfolios. Demo photos are public-domain customs seizure photographs and a CC BY-SA shop window. Team: Bernard Liu and Timotej Cvikl, Vanta Point Labs. https://meta.vantapointlabs.com/