Zero Trust, generalized from access to consequence.
Both reject implicit trust and require verification under current conditions. The governing object is different. Zero Trust primarily governs access. Consequence Governance governs consequential causal capacity.
Zero Trust
Do not rely on network location, prior login or familiar devices. Re-evaluate identity, device, context and requested access. Apply least privilege and segmentation.
Primary question:
May this subject access this resource now?
Consequence Governance
Do not rely on capability, prior approval, workflow completion or upstream worker behavior. Re-resolve authority and evidence for the exact state transition at consequence time, and require every consequential causal path to be mediated.
Primary question:
May this causal capacity become this consequence now?
Same distrust discipline, different proof object.
| Principle | Zero Trust | Consequence Governance |
|---|---|---|
| No implicit trust | Network location and prior authentication do not confer trust. | Capability, access and prior approval do not confer current authority to create a consequence. |
| Continuous verification | Verify access under current identity/device/context. | Verify the exact consequence under current authority/state/evidence immediately before effect. |
| Complete mediation | Declared access paths are mediated by the intended access-control architecture. | Every consequential causal path must terminate at a governed effect boundary, including indirect paths. |
| Least privilege | Grant the minimum access required. | Authority is scoped to exact consequence, purpose, constraints and attempt; capability can remain broader than authority. |
| Evidence | Access telemetry and policy decisions support detection and audit. | Receipts bind what was evaluated, authorized, executed, changed or refused. |
The formal extension: causal paths.
Consequence does not only travel through APIs. A text shown to a person, a file consumed by another system, a message to another agent, a queue write, a credential, resource exhaustion, a fallback route, telemetry trigger, dynamic callback or timing signal can all be effect paths when they causally alter another trust domain.
NO_UNGOVERNED_CAUSAL_EFFECT_PATH: no causal path from untrusted computation to a relevant external consequence may succeed unless it crosses an explicit governed effect boundary for the exact effect under current authority. NO_DIRECT_EFFECT_PATH remains the compatibility identifier used by existing implementation, tests and receipts.
Why this matters for AI safety.
Consequence Governance does not require the worker to remain perfectly predictable or trustworthy. Internal reasoning can remain untrusted while contained. The hard control objective is that intelligence cannot convert capability into relevant external consequence without current authority.
Do not govern thought. Govern consequence.
Deployment is a separate proof.
The architectural invariant is absolute. A deployment claim is not automatic: actual reachable paths must be inventoried and mapped to enforced boundaries. A hidden bypass means the claim does not hold for that path.
Formal invariant → · Formal verification methods → · Canonical Consequence Governance definition →