One search. Every version retained.
| Task | Full task score | Non-target preservation |
|---|
Preservation was a reporting metric in this historical contract, not a retrospective promotion gate. Declared mutation actions do not establish actual structural changes.
Inspect the original candidate record
Optimizer usage, accounted separately
Tool implementations, Schemas, and prompt contributions changed; cordis.patch did not. This run does not establish topology-search benefits. Calls count DSH adapter streams; provider-internal retries are not inferred.
Independent validation on new tasks
| Arms | Mean | Scored | Execution failed | Cancelled |
|---|
The denominator includes every registered task. Business performance, model variability, and transport failures all affect these results; the entire mean difference cannot be attributed to plugin quality.
Task-by-task, repetition-by-repetition comparison · including failures and cancellations
| Task / Repetition | Baseline | Mutation | Candidate status | Delta |
|---|
Keep the capabilities. Keep the evidence.
Non-promoted candidates can be exported as research assets for inspection and further mutation. Imports preserve byte identity; historical scores do not enter new selection decisions.
- Archived candidate
- Package SHA256
- Research archive SHA256
From understanding to hands-on use
This replay presents completed historical evidence. Configure a model and follow the documented commands to evolve with your own tasks and evaluation contract. Capability installation can be tried independently.
This page does not offer an unverified best-plugin download. Historical non-promoted packages are not recommended defaults.