Technical glossary
The language of governed AI execution.
Clear terminology prevents security controls, heuristic scores, calibrated probabilities, and marketing claims from being treated as interchangeable.
AI gateway
A control layer between applications and AI providers that centralizes authentication, policy, routing, rate limits, observability, cost controls, and integrations.
Bounded repair
A reason-coded correction attempt with explicit limits, followed by validation. Bounded repair prevents an agent from looping indefinitely or quietly changing the acceptance criteria.
Weakest-link gate
An acceptance rule in which the minimum critical-channel result controls release. It is useful when success in one area cannot compensate for failure in another.
Provenance
The evidence needed to reconstruct how a result was produced, including identities, policy, versions, providers, events, validations, repairs, approvals, and artifact references.