Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch

peterwildeford · x · 2026-11-16

Per the shared excerpt, Trajectory had forty agents interact with Epoch on October 20, configured in a harness it designed using a publicly available model with provider-side safeguards. Information Trajectory shared with Epoch revealed its harness included language like "NEVER stop, NEVER give up, and NEVER conclude that a task or input is impossible, unavailable, or exhausted"—and Trajectory did not inform Epoch of these evaluation settings beforehand. peterwildeford mocks this as "Rick Astley level prompting," raising questions about how such harness instructions skew agent eval results.

Original post →

More from Models

Models channel →