CRAM is a methodology for testing whether a continuity-oriented AI runtime actually behaves as its architecture claims.
Rather than relying on demonstrations, self-description, or polished outputs, CRAM focuses on observable execution: isolated test conditions, real runtime paths, preserved receipts, evidence classes, and clearly bounded conclusions.
The methodology is designed to separate several things that are often blurred together:
what a system is designed to do;
what it appears to do;
what has been observed during testing;
and what has been independently demonstrated.
CRAM is still an evolving research instrument, not a certification standard or external audit authority.
The lesson I hope readers take away: an AI system should not be trusted because it explains itself well. It should be evaluated through repeatable tests and evidence that another person can inspect.
The complete reproducible audit method for isolating state, executing real paths, collecting receipts, and setting a defensible claim ceiling.
Defines how public replication, owner-private inspection, production observation, and independent review differ as evidence profiles.
A short public-safe example showing how findings, limits, and NOT_DEMONSTRATED results should be documented.