Xiaolei Wang’s ABTdomain experiment separates brand-name recognition from evidence of how a domain is used. Jev answers bounded questions; code routes cases to Release, Watch or Flag, and a reviewer remains responsible for the eventual judgment.

I read the complete article. I did not rerun its API requests or verify its domain records. All numerical results below are ABTdomain’s observations, not independently reproduced measurements.

Ask less, collect more

  • The first layer sees the brand description and only the part of the domain before its suffix; it asks whether the name refers to the brand.
  • Ordinary DNS and HTTP checks send inactive domains to Watch without a model call.
  • Active cases receive a narrower use question, with registration, DNS, page-title, metadata and text-excerpt evidence—not raw HTML.
  • Flags are review referrals, not findings that a domain infringes a trademark.

The reported pilot

  • ABTdomain tested 100 Apple-containing domains, including hand-added controls, using Jev 1.13.0.
  • Its name stage released 18, flagged 24 and passed 58 onward; 26 of those were active enough for the use stage.
  • The use stage released 13, watched 12 and flagged one. These are routing counts, not accuracy scores.
  • The configured cutoffs were 0.3 and 0.7, with means of repeated calls in the illustrated cases; no independently labeled validation establishes those thresholds.

Two revealing failure boundaries

  • The author reports that applepie.top and applepie.net received 0.39 and 0.29 when suffixes were included—enough to cross a routing boundary. Removing suffixes makes that stage deliberately invariant to them.
  • A repair-services site went from a reported 0.69 on its name to 0.80 with use evidence. That is a changed model judgment, not proof of unlawful behavior.
  • Describing a synthetic community as for-profit raised its reported score, but the experiment does not establish the truth of registration intent or legal bad faith.

The author calls this a workflow plan, not a system for real judgments. It omits visual evidence, a dispute-case database and later case processing; its question also excludes some situations covered by the Uniform Domain-Name Dispute-Resolution Policy.

The reusable contribution is the decomposition: isolate evidence and failure modes by stage. A typed probability is still neither a correct decision nor a legal conclusion.