AI autonomy × Epistemic integrity

Every override of the AI's advice is a documented, blameable act in a verifiable record — so humans stop overriding. Advisory on paper, delegated in practice, with humans holding the blame and the machine holding the steering. The second-order effect: whatever biases sit in the training data become the unalterable policy limits of the institution, because no official will take the liability of contradicting them. Status migrates toward unaudited discretion — the people still allowed to make unjustifiable gut calls — while everyone else rubber-stamps.

Provenance becomes the chokepoint. When every agent action is logged and verifiable, the bottleneck shifts from capability to the right to be trusted — whoever owns the attestation layer becomes a sovereign. And agents freeze: perfect attribution means perfect liability, so insurers and courts make delegation paralytically cautious. Humans reorganise around being the legally responsible body — professional signatories with no hand near the wheel. The cleanliness is the cage. It avoids runaway automation crises; it is also too slow to capitalise on anything.

People don't trust their AI advisor because it's accurate — everything around it is fog. They trust it because it's theirs, consistent, and the only entity whose outputs they can replay. The result is pathological monogamy — one advisor per person, its drift becoming their entire worldview. And when digital records can't be verified, distance itself becomes unpriceable: insurance and remote lending collapse, and contract enforcement, credit, and security re-anchor on the one thing that can't be faked — physical custody and face-to-face attestation. The fog doesn't end trust; it drags it back within arm's reach.

Agents can't navigate a fake public space, so they retract into sealed trust enclaves — corporate, cartel, family office — transacting at machine speed inside the walls while cross-enclave commerce reverts to handshakes and physical audits. The attack surface shifts with them: why deceive voters when you can poison the sensor data of the one procurement agent that moves a supply chain? Disinformation becomes precision B2B targeting of machines rather than populations. The dark version is warring enclosures; the redeeming one is that physical accountability re-emerges exactly where digital trust died.