The agent found every page certified unreachable. Links cut search cost and keep working when search is off; they do not change what it reaches

A wrong note that passed review spread only where the agent could have checked it, and only on Haiku 4.5: 16 of 24, against 0 of 24 on Sonnet 5 and Opus 5

The harness behind two DOI-archived benchmark reports, and why a study premise can be killed in one working day

A benchmark that killed its own hypothesis: strong models re-derive what they can check, and a weak model trusts a stale note over the evidence in front of it

A reproducible benchmark: 42.7% to 98.7% on knowledge-trap questions, and a platform that learns from empty