Owned public project · rudycelekli/gradia-reward-loop
gradia-reward-loop · Tracing reward hacking into an RL diagnostic
Problem
A scorer exploit can matter differently once it becomes a training signal.
Design and implementation
Gradia Reward Loop carries oracle-witnessed pass/wrong pairs into a training-time experiment and records evidence bundles for replay. The project describes a paired GRPO diagnostic with single-variable localization.
Evidence and scope
The pinned source supports the documented diagnostic design. A broader effect on policy training needs independent replication and holdouts.