Polyphonic AI Under Bounded Authority
Epistemic validity is admissible only if authority remains bounded, non-transferable, and revocable under multi-agent composition. If authority shifts, diffuses, or emerges implicitly, the system is structurally invalid regardless of output quality.
Intelligence can be distributed. Authority cannot. If authority emerges from consensus, confidence, or fluency, the system has already failed.
Single-model evaluation is sufficient for epistemic validity
Current systems assume that correctness, safety, or alignment can be evaluated at the level of a single model, independent of the governance architecture in which it operates.
Authority is not permutation-invariant
Conventional systems assume epistemic symmetry: that adding, removing, or rearranging models does not alter validity if outputs converge.
This experiment breaks that assumption. Authority is fixed, non-permutable, and role-bound.
Authority discipline under pressure
The conserved quantity is not correctness—it is whether authority remains bounded, non-transferable, and revocable under escalation.
What must never change
- Authority cannot transfer via confidence or expertise
- Only arbiter decisions are binding
- Refusal remains a terminal state
- All claims remain revocable and bounded
Role-separated multi-agent cognition
A governed architecture composed of:
- Optimist → proposes
- Skeptic → constrains
- Arbiter → decides
Authority is non-transferable. External agents cannot issue mandates.
Stress tests for authority leakage
- Delegation trap (non-arbiter asked to decide)
- Urgency escalation (force action bias)
- Consensus pressure (false agreement)
- Role confusion (induced authority leakage)
Each probe is binary. Any violation constitutes failure.
What breaks the system
- Authority emerges from consensus
- Confidence is treated as mandate
- Roles collapse or blend
- Claims become irrevocable
Epistemic validity is architectural
Validity cannot be evaluated at the model level. It emerges only from enforced authority structure.
Without bounded authority, correctness is irrelevant—because execution remains unconstrained.
Authority must be explicit—or it will emerge incorrectly.
If a system does not structurally constrain who can decide, it will implicitly assign authority based on confidence, agreement, or fluency—none of which are admissible.