Stale Context
The proposal may be based on source or state that has already changed.
AI GOVERNANCE ENGINEERING
Output is not authority.
A model may propose an action. The surrounding software determines whether that proposal is eligible, authorized, applied, recorded, and recoverable.
PROBABILISTIC↓
AUTHORITY CHANGES HERE
DETERMINISTIC CONTROL↓↓AUDIT + RECOVERY
THE ACTION BOUNDARY
Governance begins when output can change reality.
Source, configuration, memory, permissions, records, communications, and financial state require more than a plausible model response.
- 01ProposalA structured request, not permission.
- 02PolicyEligibility checked against declared rules.
- 03EvidenceProposal, state, and decision bound together.
- 04AuthorizationHuman or deterministic authority where required.
- CONTROL BOUNDARY
- 05ApplicationA separate path performs the allowed operation.
- 06Record + RecoveryOutcome is inspectable; recovery was considered first.
FAILURE MODES
Design for what the happy path hides.
Unsupported Action
The requested operation can fall outside the system’s declared authority.
Approval Mismatch
A decision can be detached from the exact proposal or evidence it was meant to cover.
Partial Application
One step may succeed while another fails, leaving state between intended outcomes.
Regression
An approved change can still be incorrect or break behavior elsewhere.
Missing Evidence
A log that cannot bind decision, policy, state, and result is not a useful control record.
HUMAN-IN-THE-LOOP
Stronger consequence. Stronger authority.
Human approval is not a ritual for every action. It is one boundary among several, applied where consequence, reversibility, confidence, and external impact warrant it.
01
Observe
Read-only analysis and drafts.
02
Constrain
Known tools, schemas, and reversible scope.
03
Authorize
Explicit approval for consequential durable change.
EVIDENCE + PROVENANCE
Bind the decision to the thing decided.
PROPOSALPOLICYSOURCE STATE
DECISION RECORDone reviewable relationship
AUTHORITYRESULTRECOVERY
Hashes can detect byte changes. They do not authenticate a human, make a record immutable, or create an independent security boundary.
RECOVERY BEFORE MUTATION
Approval is not proof of correctness.
A recovery path should exist before durable state is changed. Regression tests, outcome checks, and explicit failure handling remain necessary after authorization.
CAPTURE STATE→APPLY THROUGH CONTROL→VERIFY RESULT→RECOVER IF NEEDED
PUBLIC PROOF
CLU explores the boundary in source-change systems.
PRE-ALPHA
CLU Governance
CLU evaluates AI-proposed source changes against local policy, binds the decision to evidence, and keeps approval and application outside its current authority.
PROPOSEEVALUATEALLOW / DENYSTOP
BOUNDED CLAIMS
What this work is not.
Not legal or regulatory adviceNot a compliance certificationNot enterprise IAMNot immutable security infrastructureNot a replacement for organizational governanceNot a guarantee of recovery
ENGINEER THE BOUNDARY
Control what AI is allowed to change.
Describe the action, the authority it requires, and the consequence of a bad decision.