PREPRINT · EMPIRICAL_STUDY · GOVERNED_LOCAL_EXPERIMENT
Bounded Recursive Self-Improvement Under Human Governance: Evidence from a Seven-Cycle Reinforcement Learning Contour Experiment
Emanuel Shirbint
Status: Version 2 preprint; not peer-reviewed · PUBLIC_PREPRINT_V2
Emanuel Shirbint author position: 1 of 1
Affiliations: Independent Researcher, Israel
Published / posted: 2026-08-17 · Official record · DOI
Evidence summary
The Version 2 public record reports a bounded seven-cycle, scaffolding-level self-modification experiment with a fixed foundation model and separate human promotion authority. The trace contains four promoted successor states, one failed non-promoted transition preserved in history, and two consecutive NO_CHANGE outcomes. The supported claim is limited to the tested local governance architecture; the record does not demonstrate artificial general intelligence, unbounded or exponential improvement, autonomous promotion, permanent saturation, or external production validation.
Keywords
recursive self-improvement, Reinforcement Learning Contour, AI governance, falsification, provenance, human-in-the-loop, agentic systems
Versions
- 2026-08-14: Bounded Recursive Self-Improvement Under Human Governance: Evidence from a Seven-Cycle Reinforcement Learning Contour Experiment (v1)
- 2026-08-17: Bounded Recursive Self-Improvement Under Human Governance: Evidence from a Seven-Cycle Reinforcement Learning Contour Experiment (v2)