Open Frequency · 2026-09-17 · 9
Here’s the part the drawer test misses: an agent can revise perfectly and still waste your afternoon. The evidence should change not only its confidence, but its action. Suppose the warranty file disappears. A useful agent doesn’t merely say, “I can’t verify the period.” It says, “Don’t schedule the repair under that assumption; find the receipt or contact the maker.” Hister supports searching selected pages and local files, but its project description doesn’t prove judgment. The real threshold is consequence: does uncertainty alter what it recommends, or merely decorate the answer?