Known Answer tests show AI self-evaluation isn't proof

The experiments build on the author's earlier Self Evaluation work, where an AI was asked to judge its own or another AI's answers. That raised the question of what standard could confirm such a judgment, since the evaluator is itself an AI. PI-048 and PI-049 introduce a Known Answer prepared outside the AI, so a response like 4 against the known 3 exposes an error even when the AI calls the answer correct.

PI-050 extends the check to the question itself: if a prompt states a false premise, the AI may accept it and then rate its own answer as correct, so the premise needs independent checking. PI-051 looks at Consistency across repeated questions, but the author stresses that stability is not the same as correctness — an AI can be consistently wrong. The author frames AI evaluation as one layer among several, to be combined with rules, tests, external evidence, and human review where needed.