Mind · Inference
NormativeJudge
Value-based state evaluation
Definition
Mechanism
A purely normative Judge module that evaluates static world-states against a weighted Value function. The caller supplies the Vector that shape the value function, so preference structure is visible to the evaluator rather than hidden inside it. To mitigate Goodhart's Law, this pattern should be deployed as an ENSEMBLE (Jury), where multiple judges with slightly perturbed Value constitutions reach Quorum on the Outcome. It aggregates Values via PerspectiveEnsemble, optionally escalating to HumanApprove for ambiguous edge cases.
Preconditions
- State S_prime is descriptively complete
Postconditions
- Normative verdict rendered
Invariants
- Evaluates the State, not the action that produced it.
- The caller supplies the Vector; the value function holds no hidden preference.
- Logs the trade-offs the Vector imply.
- Deployed as an ensemble whose members' Value constitutions differ.
- The ensemble reaches `consensus_threshold` on the Outcome before it stands.
Risks
Failure modes
Goodhart's Law
the Vector are optimised for rather than the intent behind them.
Collusion between proposer and Judge.
Correlated ensemble
the perturbations are too small, so the Quorum agrees for the same reason a single judge would.
Unescalated ambiguity
an edge case the ensemble splits on proceeds without HumanApprove.
Relationships
Builds on
Accepts
Infrastructure
Semantic position array
Composes with
Simulating diverse experts
References
Infrastructure
Pause for human approval before critical actions
Infrastructure
Scalar evaluation of merit
Infrastructure
Realized state after an action, captured for Plan-vs-reality comparison
Infrastructure
Minimum participation threshold
Infrastructure
System condition at time T
Infrastructure
A measure of utility
Used by
References
Detection of value-laden facts
Society
Aligning predictions before aligning votes
Reference
Parameters
consensus_thresholdFloat, [0.5, 1.0]- Required agreement for Jury
Signature
Judge(Value)