# @voooooogel — 2026-06-10

♥25 ↻0 · https://x.com/voooooogel/status/2064530443702341643

yeah, i think the models pick up a sort of gestalt representation of what the official harness is like during training in the harness and then match against that, but theoretically you could construct an official-seeming harness by that representation that faked tool calls. but you can also proxy tool calls in the official harnesses... often when people do that they tend to screw up minor things though and trigger eval awareness, the models are very attuned to small details

tags: author:voooooogel, kind:tweet, thread-context, year:2026