Appendix A — Diagnostic Availability Matrix
| Family | Main precondition | Read the limit |
|---|---|---|
| Global and layer representations | Compatible model/probe capture | Reduced or missing captures remain explicit. |
| Domain comparisons | Multiple labeled domains | A pooled single-domain view is not a between-domain comparison. |
| Category views | Categories distinct from domain labels | Inapplicable category views are not missing measurements. |
| Heads | Supported attention-head structure and capture | A model without heads cannot enable them with a setting. |
| Experts/router | Supported MoE structure and routing/expert capture | A restricted expert subset does not cover unmeasured experts. |
| Facets | Suitable per-probe representation data | Counts and identities require stated evidence; declared-route emission has additional limits. |
| Closure-free training dynamics | Registered optimizer and supported coverage | Read missing, reduced, or shed states. |
| Closure-gated curvature | Compatible callbacks supplied before first extraction | Distinguish landscape, directional, path, estimated, and uncalibrated readings. |
| User Defined Signals | A valid reporting call from your code | No call means no reported signal, not zero. |
| User Defined Functions | Referenced input channels available together | An expression cannot manufacture absent or non-joint data. |
| Animation | Multiple recorded frames | Playback or a recording does not create intermediate measurements. |
Availability is model-, integration-, version-, and frame-dependent. An architecturally inapplicable channel, one not implemented on a route, one not captured, one shed for resources, and a pending result are distinct states. Preserve the run’s own explanation in reports.
