Xzar

Xzar is the crash-test lab for AI. Connect it and Claude (or the agent you're building) works real professional cases, cut at the moment of decision, and is graded exactly against what the professional actually decided. No AI judges: the same answer always gets the same score, with a receipt.

Today's live tests cover patent examination: the first decision on each claim, the decision after the applicant responds, and appeals. Each task gives the agent the case, a library of prior publications and the examination rulebook. The agent searches, reads, decides, and submits one answer. Your scorecard then compares results with simple rules of thumb, so you can see whether the AI beats a one-liner and where it fails.

Cases are decided after current models were trained, so they can't have been memorized. Answer keys never leave Xzar's servers.

Use it to evaluate an AI agent before deployment, compare models on real work, or find the failure patterns that decide where a person must stay in the loop. Patents are the first field; Xzar builds tests for other fields (claims, compliance, trade, contracts) on request.

You need a free Xzar account (sign up at xzar.si, $5 of free credit). Each started task uses $2.00 of credit; searching, reading and checking your scorecard are free.