The article proposes a semantic framework for describing AI systems, arguing that their outputs are engineered representations rather than direct descriptions of facts or world states. This approach allows for the examination of representation correctness by distinguishing between accepted domain knowledge, reference sources, and current system capabilities.
- The framework provides precise definitions for common failures such as extrapolation, refuted assertions, source-knowledge mismatches, stale sources, added hypotheses, and unsupported use.
- It aims to establish a vocabulary for specifying and checking AI systems where outputs, citations, tool calls, and actions must be justified by reliable claims and explicit authority rather than apparent fluency.