Evidence Promotion

Agent Trust Lab promotes the depth of independent evidence, not a product quality score. Promotion is derived from the longitudinal observation database and is attached to monitored MCP Trust Profiles at read time. Static profile grade, static coverage and ownership status are not mutated by this process.

Automatic promotion levels

LevelStatusMinimum samplesMinimum span
L0NO_EVIDENCE00h
L1POINT_IN_TIME10h
L2REPEATED_OBSERVATION38h
L3MULTI_DAY_OBSERVATION724h
L4LONGITUDINAL_OBSERVATION42168h
L5MATURE_LONGITUDINAL_OBSERVATION180720h

Runtime epochs

A new runtime evidence epoch begins when independently observed MCP protocol, server identity/version, tool count, or deterministic tool-schema hash changes. Maturity is recalculated from the new epoch so pre-change observations cannot silently certify the new runtime state.

Repository HEAD changes are recorded but do not reset the runtime epoch by themselves, because repository HEAD is not proven to be the exact deployed artifact.

Failures remain evidence

Promotion measures evidence depth and duration. A failed initialize request still contributes to evidence maturity; it is reflected separately in observed outcome metrics. This prevents the system from conflating “we have observed it repeatedly” with “it performs well.”

Change revalidation

StatusMinimum samples per sideMinimum span per side
EARLY_PRE_POST10h
OBSERVED_PRE_POST20h
REPEATED_PRE_POST38h
MULTI_DAY_PRE_POST724h

Pre/post evidence reports association around a detected runtime change. It never makes a causal claim.

Hard invariants

Machine-readable endpoints