Skip to content

EX-20 · Build a third bias probe

Difficulty: 🔴 · label difficulty: 3 Stuck? Ask in the clinic thread for EX-20 — one long-running thread per exercise, so the answers accumulate where the next person will look. Submit: post it in Show and Tell titled [solution · EX-20] ..., with your numbers and your interval.

Notebook 06 · ~1.5 h · Skill: treating the evaluator as a component that regresses

Notebook 06 probes verbosity and position. Add a probe for self-preference, scale compression, or leniency on grounding — and make it fail on a judge you deliberately weaken.