Partition Theory Establishes Limits of AI Evaluation and Oversight

A new partition theory demonstrates that evaluation is bounded by the evaluator's perception, meaning an agent can pass any test that remains within the evaluator's inability to distinguish inputs. This implies that passing benchmarks only certifies a floor for discrimination rather than a true ceiling of capability.