x

LESSWRONG

LW

adlnk — LessWrong

adlnk

adlnk

Message

I do science.

18h

adlnk

I do science.

Improving Audit Realism With Inference Time Compute and Deployment Scaffolds

by Axel Ahlqvist, Richard Guan, Juan Pablo Rivera, adlnk, mitroitskii, alexandrasouly, RobertKirk, and John Hughes

tl;dr: We introduce critique refinement, a method to iteratively improve the realism of auditor outputs in Petri, and DISH (Deployment Imitating SWE-agent Harness), a method to audit a model within a coding agent scaffold to ensure the model receives a real system prompt, tools and scaffold injections. These techniques compose...