SAGO measures generalization as per-example stability across prompt variations

SAGO evaluates whether an LLM preserves behavior when the same input is expressed in semantically equivalent ways. It combines multiple datasets and variation families with output, activation, confidence and mirroring measurements.