Bounded Recursive Self-Improvement Under Human Governance: Evidence from a Seven-Cycle Reinforcement Learning Contour Experiment

<jats:p>Recursive self-improvement is commonly framed as repeated self-modification that yields progressively greater capability. That framing leaves a prior methodological problem unresolved: how can a system distinguish a genuine improvement from a change that merely satisfies an evaluation procedure that the same system helped construct? This paper reports a seven-cycle instrumented experiment using a governed implementation of the Reinforcement Learning Contour (RLC), a framework previously proposed for traceable transformation of the mechanisms governing future inquiry rather than optimization of a scalar reward. A cryptographically frozen baseline, RLC0, was subjected to bounded cycles of observation, measurement, diagnosis, preregistration, sandboxed candidate construction, evaluation, decision, post-cycle integrity checking, and separate human promotion authorization. Candidate changes were prohibited from directly modifying their parent states, the Sigma-sensitive governance layer, authoritative external systems, or historical evidence. The seven-cycle trace produced four promoted states (RLC1-RLC4), one failed non-promoted transition, a prospective revision of the integrity boundary, and two consecutive independent NO_CHANGE outcomes. The failed third cycle was retained rather than rewritten; forensic analysis identified a mismatch between immutable-reference assumptions and volatile operational artifacts, leading to Boundary v1.1. Subsequent cycles resumed from the last valid parent and added boundary-aware inheritance validation and deterministic resource/effect evidence reconciliation. Cycles 6 and 7 found no new qualifying deficiency within the bounded search protocol and created no successor state. The experiment does not demonstrate artificial general intelligence, unbounded or exponential recursive growth, or permanent saturation. It supports a narrower claim: under the tested conditions, a human-governed recursive architecture can accumulate several validated operational capabilities while preserving failure, revising measurement boundaries prospectively, preventing automatic promotion, and terminating repeated searches without manufacturing an improvement when the evidence does not support one.</jats:p>