Definition
The level of the stack at which an evaluation holds: model, agent, delegation, institution, or consequence. Every suite and case carries its layer so a clean result at one level is never read as a result at another.
It matters because assurance overstates itself when a result at one layer is read as a result at another. A model eval says nothing about whether the authority grant is still valid or whether the institution retains the correction capacity to act on what the eval found. A layer also has to be staffed by someone other than the executor, which is what evaluation independence names.
MeasurableVisibility
Tests that show it is in place
- Every eval case declares its layer, and the declaration is enforced by the suite schema.
- Reports state the layer alongside the result rather than in an appendix.
- A claim about institution-level behavior cites at least one institution-layer case.
Evidence to keep
- Artifact: An eval suite schema with a required layer field on every case.
- Behavior: A claim rejected in review for citing a model-layer result as a delegation-layer finding.
- Metric: Distribution of cases across the five layers per suite.
References
- Evals index · Reference