We pulled our own numbers.
An internal audit found that every benchmark run we had published was cut short before the agents finished investigating — so the scores were floors, not measurements. They are withdrawn until we have a run that isn't truncated, and what replaces them will be generated from that run's own artifacts rather than written by hand. A tool that refuses to report an unproven finding should hold itself to the same rule.
In the run we had been quoting, the coordinating agent finished while five subagents were still working. They were given sixty seconds and then cancelled mid-request. One had just had a finding rejected over a formatting error and never got to re-file it. The score that run produced was real, but it measured an investigation that stopped early — not what the tool can find.
Agents now get as long as their own turn and cost budgets allow, every run records whether anything was cut off, and the benchmark harness refuses to score a run it detects as truncated rather than publishing the number anyway.