Link to current text
C.32.ACE:3 - Forces
| Force | Tension |
|---|
| Variant learning | Candidate architectures may be valuable even when they lose the selection being made. |
| Fair comparison | Eval results are useful only when context, budgets, windows, units (when required by the scale), and missing-data treatment are explicit. |
| Trade-off pressure | Improving one architecture characteristic can worsen another. |
| Automation value | Frequent automated evals reveal drift early, but their results can be overread. |
| Error prevention | Some eval operations are tests, yet error checking must not replace variant comparison. |
| Evolution | A useful eval can expire when the source-currentness relation, environment, declared holon-level ref, or scale window changes. |