Literature review. LLM judges, learned verifiers, plan verifiers, runtime guardrails and agents gaming their evaluators — how each works, and where MORPHEUS differs.
Literature review. Routers that predict from the prompt rarely generalize; cascades that check the cheap answer depend on the signal — and where Rung differs.
Predicting the best model from the prompt recovered none of the available gain; climbing a price ladder on evidence did. The check that says “escalate” is the hard part.
Two small RL experiments. A reward that cannot check a refusal cannot teach one; a strong model was already honest, but a third of its proofs contained a false step.
Nine graders, one frozen set. The grader with the highest agreement accepted half of the planted errors; the only graders that accepted none had some of the lowest agreement.