| --- |
| license: cc-by-4.0 |
| language: |
| - en |
| pretty_name: Governed Skill Evolution from Persistent Agent Experience |
| size_categories: |
| - n<1K |
| tags: |
| - ai-agents |
| - agent-skills |
| - persistent-memory |
| - ablation |
| - cross-model-transfer |
| - reproducibility |
| configs: |
| - config_name: replicate_results |
| data_files: |
| - split: train |
| path: data/replicate_results.jsonl |
| - config_name: aggregate_results |
| data_files: |
| - split: train |
| path: data/aggregate_results.jsonl |
| - config_name: skill_impact_ledger |
| data_files: |
| - split: train |
| path: data/skill_impact_ledger.jsonl |
| --- |
| |
| # Governed Skill Evolution from Persistent Agent Experience |
|
|
| Prospective ablation and cross-model transfer study of three experience-retention conditions for governed Agent Skill evolution: no persistent history, flat chronological history, and a persistent Pattern Registry with a forward-chained Skill Impact Ledger. |
|
|
| - **Author:** Song Luo |
| - **Version:** 1.0.0 |
| - **Source snapshot:** [`d717c32396cfff1bef2800296541a70e9b4cabb8`](https://github.com/rrrrrredy/governed-skill-evolution/commit/d717c32396cfff1bef2800296541a70e9b4cabb8) |
| - **Canonical repository:** [rrrrrredy/governed-skill-evolution](https://github.com/rrrrrredy/governed-skill-evolution) |
| - **Zenodo:** [10.5281/zenodo.22240411](https://doi.org/10.5281/zenodo.22240411) |
| - **Full study harness:** [Runtime Evolution Workbench](https://github.com/rrrrrredy/runtime-evolution-workbench/tree/main/research/wiki-evolution-study) |
|
|
| ## Viewer subsets |
|
|
| - `replicate_results`: 9 rows, one per condition and replicate. |
| - `aggregate_results`: 3 condition-level descriptive summaries. |
| - `skill_impact_ledger`: 36 compact ledger entries. The full verified ledger, including digest material, remains under `results/`. |
|
|
| ## Main result |
|
|
| Persistent Wiki did not outperform flat history on task quality in this setting. It used fewer input tokens and produced a larger target-model Skill gain, but had lower final task quality and more rollbacks. The release therefore preserves negative and mixed results rather than treating persistent state as automatically beneficial. |
|
|
| ## Evidence boundary |
|
|
| This is prospective descriptive evidence for one synthetic deterministic grader, three replicate sequences, and two Codex model versions in one model family. It is not a causal provider comparison, safety certification, or broad claim about Agent performance. |
|
|
| Source code and machine-readable protocols are Apache-2.0. The manuscript, documentation, result tables, figures, and release metadata are CC BY 4.0; see `CONTENT-LICENSE.md`. |
|
|
