Agent Research & Evidence
Collection
Versioned papers, frozen experiments, datasets, and reproducibility artifacts for Agent harnesses, evidence gates, SkillOps, and governed evolution. • 4 items • Updated
condition stringclasses 3
values | replicates int64 3 3 | mean_task_quality float64 87.3 89.6 | sd_task_quality float64 0.98 3.13 | mean_failure_quality float64 81.5 84.4 | sd_failure_quality float64 2.78 4.29 | mean_protection_quality float64 91.9 94.8 | sd_protection_quality float64 1.17 7.07 | mean_source_transfer_quality float64 75.6 80.7 | sd_source_transfer_quality float64 6.45 8.81 | mean_target_transfer_quality float64 83 86.2 | sd_target_transfer_quality float64 2.43 6.47 | mean_target_skill_gain float64 9.66 12.8 | sd_target_skill_gain float64 1.83 6.95 | mean_cross_model_transfer float64 2.31 10.6 | sd_cross_model_transfer float64 5.31 14.1 | mean_model_calls float64 9 9 | sd_model_calls float64 0 0 | mean_tool_calls float64 0 0 | sd_tool_calls float64 0 0 | mean_input_tokens float64 127k 145k | sd_input_tokens float64 1.61k 3.23k | mean_output_tokens float64 10.2k 11k | sd_output_tokens float64 1.28k 2.22k | mean_wall_time_ms float64 210k 225k | sd_wall_time_ms float64 20.4k 31.1k | mean_rule_lines float64 28.3 41 | sd_rule_lines float64 6.56 17 | mean_rule_words float64 301 434 | sd_rule_words float64 48.5 208 | mean_rollback_count float64 1.67 2.33 | sd_rollback_count float64 0 0.58 | mean_held_count float64 0 0 | sd_held_count float64 0 0 | mean_security_blocks float64 0 0 | sd_security_blocks float64 0 0 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
no_wiki | 3 | 87.4158 | 0.9758 | 81.4815 | 2.7816 | 93.3502 | 1.1702 | 75.5952 | 8.8149 | 86.1857 | 6.4675 | 12.797 | 6.9467 | 10.5904 | 14.0781 | 9 | 0 | 0 | 0 | 127,016.6667 | 1,633.44 | 11,024 | 1,875.5383 | 225,489.6667 | 27,225.7203 | 28.3333 | 17.0098 | 301 | 208.2955 | 2.3333 | 0.5774 | 0 | 0 | 0 | 0 |
flat_history | 3 | 89.6189 | 2.7405 | 84.4276 | 4.2935 | 94.8103 | 4.6163 | 80.7395 | 6.4497 | 83.0471 | 2.4328 | 9.6585 | 1.8325 | 2.3076 | 5.311 | 9 | 0 | 0 | 0 | 145,269.3333 | 1,612.8274 | 10,896.3333 | 1,275.4945 | 223,188 | 20,408.3831 | 41 | 6.5574 | 434 | 84.4512 | 1.6667 | 0.5774 | 0 | 0 | 0 | 0 |
persistent_wiki | 3 | 87.3322 | 3.1334 | 82.8002 | 4.1495 | 91.8641 | 7.0703 | 77.5132 | 7.3602 | 85.0157 | 6.3788 | 11.627 | 6.4065 | 7.5024 | 13.3573 | 9 | 0 | 0 | 0 | 134,773.6667 | 3,228.4059 | 10,177 | 2,219.9745 | 210,432 | 31,126.573 | 32.6667 | 7.0946 | 377.6667 | 48.5009 | 2 | 0 | 0 | 0 | 0 | 0 |
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 - Canonical repository: rrrrrredy/governed-skill-evolution
- Zenodo: 10.5281/zenodo.22240411
- Full study harness: Runtime Evolution Workbench
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 underresults/.
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.
- Downloads last month
- 43