Governance
LLMO governance defines the constraints within which machine-mediated work operates. It is not compliance theater. It is the structural boundary between useful automation and uncontrolled output.Trust boundaries
Every decision in an LLMO system operates within a defined trust boundary:- Domain scope: what topics or entity types the system is authorized to address
- Risk tier: the consequence level of errors in this domain (low, medium, high, critical)
- Error bound: the maximum acceptable error rate for verified decisions
- Latency constraint: the time window within which a decision must be produced
- Authority level: who or what is permitted to make, approve, or override decisions
Policy constraints
Policies define what the system can and cannot do within its trust boundaries:- Permitted actions: what outputs the system may produce
- Prohibited actions: what outputs are never acceptable regardless of model confidence
- Conditional actions: outputs that require human approval before delivery
- Escalation triggers: conditions that force routing to a human reviewer
- Refusal conditions: states where the system must decline rather than guess
Audit requirements
Every verified decision must produce an audit trail that includes:- The input that initiated the decision
- The sources consulted
- The claims used in reasoning
- The model outputs generated
- The evaluation results
- The policy checks applied
- Whether human review occurred and what the outcome was
- The final decision and its confidence level
- Timestamps for every state transition
Approval classes
Not all decisions carry equal weight. LLMO governance defines approval classes:| Class | Description | Approval requirement |
|---|---|---|
| Automated | Low-risk, well-calibrated, high-confidence | Harness approves, logged |
| Reviewed | Medium-risk or edge-case | Human reviews before delivery |
| Escalated | High-risk, ambiguous, or novel | Senior domain expert approves |
| Refused | Insufficient truth or out-of-bounds | System declines, logged |
Escalation rules
Escalation is not failure. It is the system operating correctly under uncertainty. Escalation occurs when:- Model confidence falls below the threshold for the current risk tier
- Multiple sources conflict on a material claim
- The decision falls outside the defined domain scope
- Policy checks identify a prohibited or conditional state
- The harness detects potential calibration error above the acceptable delta
Red-team principles
LLMO systems should be adversarially tested for:- Source poisoning: can a bad actor inject false claims into the source graph?
- Prompt injection: can retrieval-augmented generation be manipulated?
- Stale data exploitation: can outdated claims be surfaced to trigger bad decisions?
- Calibration gaming: can a system be made overconfident through curated inputs?
- Audit evasion: can the logging layer be bypassed or tampered with?
Constitutional design
At the highest level, LLMO governance is constitutional. It defines not just rules but the principles from which rules are derived:- Truth integrity over output volume
- Verified decisions over fast decisions
- Refusal over confident wrongness
- Auditability over convenience
- Human accountability for high-stakes outcomes
- Bounded autonomy, never unbounded delegation