An automatically improved skill should be evaluated before being accepted.
Proposed workflow:
old skill -> evaluate
new skill -> evaluate
compare -> accept or reject
Reject improvements when:
- overall performance decreases
- safety decreases
- required deterministic checks regress
The original skill must remain available and no silent overwrite should occur.
An automatically improved skill should be evaluated before being accepted.
Proposed workflow:
old skill -> evaluate
new skill -> evaluate
compare -> accept or reject
Reject improvements when:
The original skill must remain available and no silent overwrite should occur.