Abstract
Persistent AI agents can propose changes to their own operating artifacts, but a system that both changes and judges itself creates unresolved questions of evidence custody, authority, and accountability. EvoHakiki develops an auditable control-plane architecture for evidence-governed agent evolution. Phase I is anchored exactly to release v0.10.0 of the system historically developed as Reflective Agent. It specifies evidence-integrity controls, artifact lineage, deterministic advisory policy, append-only human adjudication, auditability, and a separation between policy recommendations and authority to perform external actions.
The Phase I contribution is architectural and methodological. It does not claim demonstrated longitudinal capability improvement, validated metacognition, complete containment, or safe autonomous promotion. Phase II begins from v0.11 or an explicitly versioned successor and is planned as a longitudinal empirical evaluation of capability improvement, retention, safety regression, governance cost, and causal differences between matched experimental conditions. Preliminary observations from v0.11 or later are not treated as Phase I evidence or completed empirical results.
Availability and citation
- Recommended citationNot yet finalized
- DOINot yet assigned
- Direct PDFNot yet released
- Source and artifactsNot yet supplied
Version and correction information
- v0.10.0: Phase I architecture-and-methods anchor
- v0.11 or successor: Planned Phase II empirical starting point
No corrections have been posted because the manuscript has not been released. When a searchable-text PDF becomes available, this page is prepared to add direct PDF, DOI, citation, and Highwire Press metadata without changing its permanent URL.
This manuscript-in-preparation record intentionally uses CreativeWork, not ScholarlyArticle, and does not emit Highwire citation tags. Released-paper metadata will be enabled only after a genuine paper and searchable-text PDF exist.