In plain English
This page explains the governance layer: rules, logs, approvals, signatures, audits, permissions, and rollback tools. These controls are necessary, but they also become important failure points.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Fail-Closed Governance
In adaptive systems, errors in the control layer must not silently become approval. Fail-closed governance means uncertainty denies or pauses the action until independent review succeeds.
Fail-open hazards
Fail-open behavior appears when evaluatorA system that judges whether an AI output or candidate is acceptable. Open glossary definition outages skip tests, missing metadata defaults to trusted, hidden-test failures are ignored, signature checks are soft warnings, or canary monitors are disabled during release pressure.
Fail-closed design
A fail-closed control planeThe governance layer that decides what can run, change, access tools, or be released. Open glossary definition denies promotion when provenance is missing, halts composition when a dependency is unsigned, freezes aliases during evaluator disagreement, blocks memory consolidation during incident review, and prevents tool-permission expansion without a separate approval record.
Organizational requirement
Fail-closed controls only work if operators are allowed to tolerate delay. A culture that punishes rollbackReturning a system to an earlier known state. Open glossary definition or no-op will pressure the system toward fail-open workarounds.