A photo from the market stall in, an evidence record out.
The investigator photographs the stall, shelf or listing. SAM 3.1 isolates every logo, label and pattern. Muse Spark reads each one against the client's registered marks and scores the six global-assessment questions. The record keeps the hashes, the crops, the rationales and every model call. The lawyer decides.
1 · The photo
Recorded sweeps replay instantly. Your own photo runs live.
2 · Acting for
3 · Field notes
How it works
Segment
SAM 3.1 (Meta Model API, Responses API, streamed) is asked three short noun phrases, "brand logo", "printed brand name", "repeating pattern on a bag". Every box and lossless mask comes back with a stable object id; boxes are merged across phrases and cut from the original photo. No model picks what to look at; the segmenter finds everything that looks like a mark.
Assess
Muse Spark 1.3 sees each crop next to the office drawings of the client's registered marks and returns schema-checked JSON: what it sees, the text it reads, which client marks are candidates, six sub-scores with rationales (visual, aural, conceptual, goods, distinctiveness, imitation signals), a narrative in "highly concerning / worth checking / low concern" language, never a legal conclusion.
Record
Photo hash, crop hashes, boxes, masks, scores, rationales and the ledger of every model call (model, purpose, latency, tokens) in one JSON file per sweep, served read-only. The roll-up is a weighted mean with N/A weights redistributed, same bands as Signum, our case-assessment product: red at 65, amber at 40.
What it is not
Not a verdict. The register entries are the only comparison; the scores are research signals for the lawyer's own global assessment. The two demo clients use publicly registered marks; the demo photos are public-domain customs seizure photographs and a CC BY-SA shop window, no seller is identified.