Monitoring, recording and scoring how systems exploit seed data to produce new information and inferences.
-
Within the boundaries of an ontology, epistemic assurance records and scores how AI systems use seed information to generate new information outputs. It captures a claim’s point of origin, the nature and quality of its evidence-base, transformations that occur as information is ingested, digested, and exploited, and relevance and alignment to ontological conditions.
-
Outputs arrive without any indication of the quality of what went into them. Confidence is conveyed impressionistically, in tone and hedging, and is lost entirely when a result is passed to someone who did not produce it. The cost shows up as rework, as silent propagation of weak findings through several stages of analysis, and as decisions taken on ground that nobody has examined.
-
-
1. What makes a confidence signal usable at the point of decision, rather than ignored or over-read?
-
2. How does evidential quality degrade across a multi-step transformation, and is that degradation measurable rather than merely arguable?
-
3. What does a downstream reader need in order to rely on an assessment without repeating the work?
-
4. Can an assurance record remain interpretable outside the organisation that produced it, and what would make it portable?