Human authority
What people must remain able to see, understand, correct, govern, carry, and retire.
Persistent AI systems can accumulate state, permissions, inferences, and downstream consequences across time. This work separates three things that are often collapsed into one.
Authority ≠ Enforcement ≠ Compliance
A human may validly revoke authority over some piece of persistent state. A provider's system may or may not execute that revocation correctly. Even when execution is correct, the provider may or may not comply with the revocation downstream. A failure at one layer does not erase what happened at another: a provider's failure to execute a revocation does not erase the human authority event itself.
Where this is developed and tested
- Human-Authority Layer — the conceptual architecture (Vision v0.2 frozen) and OD-HAP-01, an executable pre-pilot protocol testing state visibility, revocation with derivative-state consequences, and explicit authority boundaries.
- The Persistent Primitive — requirements for provenance, validity, correction, authority, purpose limitation, revocation, and portability in governed continuity.
- Authority Does Not Travel by Default — what a receiving system must reevaluate before it uses relationship state transferred across a boundary.
- Governing Relationship State — lawful continuing reliance as the central governance problem.
- Human Agency Under Consequential State and its policy paper — correction, revalidation, and recoverable consequences after consequential state changes.
- When Correction Cannot Catch the System — whether human authority still reaches the reliance paths an accepted correction did not catch.
What happens after a correction is issued is examined on Consequence.