Mind · Inference
NormCheck
Detection of value-laden facts
Definition
Mechanism
A Check filter that scans the Prophet's descriptive output for normative adjectives (e.g., 'dangerous', 'unfortunate', 'unfair') that masquerade as objective facts. Detected adjectives are handled according to `action_on_detect` — marked, rewritten out, or the text rejected — preserving the separation of Is (Prophet) and Ought (Judge of Value). Utilizes Quorum, ProphetFanOut, NormativeJudge.
Preconditions
- Text claims to be descriptive
Postconditions
- Text is certified neutral, marked, or rejected.
Invariants
- Fact Preservation: under Rewrite, the rewritten text preserves every causal claim of the input.
- Neutral Tone: no text passes as certified descriptive while carrying a detected normative adjective.
Risks
Failure modes
Strips relevant safety warnings that were actually factual.
Relationships
Builds on
References
Infrastructure
Side-effect-free observational verification yielding Verified/Falsified/Unknown
Infrastructure
Scalar evaluation of merit
Value-based state evaluation
Branching future-scenario generation
Infrastructure
Minimum participation threshold
Infrastructure
A measure of utility
Nothing in this library builds on this pattern yet.
Reference
Parameters
action_on_detectEnum, {Flag, Rewrite, Reject}- Response to smuggled norms
sensitivityEnum, {Low, Medium, High, Paranoid}- Detection threshold
Signature
Check(Value)