# Reward Integrity A grader that the model cannot game. This is the real quality bar for an RL environment. > [!danger] Failure mode > The grader scores perfectly while the model learns the wrong thing. Nobody notices until the model ships. ## Practices - **Held-out graders**: keep some graders hidden from the training loop. - **Rubric versioning**: every score points to the rubric version that made it. - **Adversarial tests**: attack the verifier before the model does. - **Hack monitoring**: watch for score jumps without capability jumps. ## Alpha decay - A task stops paying once models pass it. Labs drop tasks at ~70% pass rate. - Build a pipeline that makes the *next* field. A fixed catalogue decays. ## Why it matters for Navon - An independent grader host is more credible than the vendor that sold the environment. - Signed, versioned scores fit [[RL Data Provenance]] and PQC signatures. ---- tags: #boat #deeptech links: [[RL Environment Stack]] [[RL Environment Growth Playbook]]