How to score a test
- Run every normal, incomplete, conflicting, duplicate, and sensitive case.
- Compare proposed fields and actions with the expected disposition.
- Count incorrect facts, missed exceptions, unsafe actions, and reviewer corrections.
- Measure complete handling time, including manual review and recovery.
- Record the tool, version, configuration, date, and known limitations.
What the pack can establish
It can compare configurations against a disclosed fictional dataset. It cannot establish performance on customer data, prove regulatory compliance, or predict business savings.
Publication standard
CSLM will publish benchmark results only after actually running a documented test. Any future result must include the methodology, sample size, configuration, raw synthetic cases, scoring rules, and limitations.