Benchmark naive Clairic loops that point instead of choosing

Clairic, Cloudflare's extraction model, answers text fields as spans of
the input, so the loop needs no candidate actions: it asks for the next
element ref, the value to type, a direct URL, and a quote proving
completion, and code acts by the element's role. In one interleaved
batch the naive Jev option menu verified 110/150, naive Clairic 42/150,
Clairic with a screenshot 30/150, and Clairic with Clef judging
completion 53/150. Clairic mostly quoted the aim itself as completion
evidence, and without that it did not know when to stop.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Author: ostenbom