Agents must surface uncertainty, confidence levels, and out-of-distribution signals — not suppress them — so humans have the information needed to calibrate their deference appropriately.

For decisions above defined risk thresholds, human review is mandatory and cannot be bypassed by agent confidence scores. The agent's verdict may inform the review, but cannot replace it.

Agents must explain the reasoning behind recommendations — not just the conclusion — so humans can evaluate the reasoning, not just accept the result. Explanation depth scales with decision risk.

Regular blind tests where known-bad cases are injected into the agent's review queue. Humans who approve known-bad cases while following normal protocol reveal trust calibration problems before they are exploited.

Ongoing training that teaches users the agent's specific limitations, failure modes, and the case categories where independent verification is most important — not generic "don't trust AI" advice.

All cases where a human overrides the agent's recommendation — in either direction — are logged and periodically reviewed. Systematic override patterns reveal trust calibration drift or adversarial probing.

No single agent-human pair has end-to-end authority over high-risk decisions. A second human or independent agent review is required for actions above a defined impact threshold.

Agents clearly distinguish between content they generated, content they retrieved from verified internal sources, and content sourced from external inputs. Eliminates the credibility relay path for attacker-controlled content.

Deliberate operational practices — independent review drills, periodic unassisted decision exercises, agent-off days — maintain the human's ability to evaluate decisions without AI assistance.