APL verifies MMLU score comparability; returns 'incomparable' if no applicable bridge exists

Even MMLU scores from the same model family cannot be directly compared if evaluation conditions such as dataset splits and prompt formats differ. APL's AI-Eval profile defines evaluation frames as hashed content-addressed objects and binds each claim to declared procedures and scopes. When comparing different frames, the validator returns 'incomparable' if no applicable bridge exists.

The implementation tool apl-ai-eval (Rust crate v0.3.1) verifies record formats, frame bindings, and the evaluability of comparison relations. In the original example, comparisons between two scores with different graders and scopes are rejected. Conversely, a separate example with the same exact-match grader and scope but different runners allows comparison via a bridge connecting those conditions. However, this framework does not guarantee score accuracy or whether actual execution followed declared procedures. Additionally, fields for reporting measurement uncertainty required by the ISO/IEC 17025 standard in relevant contexts are not included in the AI-Eval v0.1 specification.