Self-Growth and Controlled Self-Evolution in Intelligent Systems: Conceptual Boundaries, Continuity Protocols, and Falsifiable Evaluation
Long-lived intelligent systems must remain effective as tasks, environments, and tools change. However, self-modification, persistent capability growth, and open-ended evolution are distinct claims. This paper develops a conceptual and engineering framework that distinguishes growth proposals, persistent capability growth, controlled self-improvement, Darwinian selection, and open-ended evolution, and specifies the evidence required for each. Representative studies of skill accumulation, self-modifying agents, meta-level improvement, and continual learning support bounded autonomous improvement, but do not establish indefinite, task-universal, monotonic progress. The framework represents a system in terms of its executable configuration, capability assets, engineering identity, governance rules, event history, and logical time. It proposes a protocol linking each candidate to prior authorization, isolated experimentation, independent validation, controlled application, a stability assessment, and formal acceptance. Configuration restoration is treated as a forward state transition that preserves the historical record. A falsifiable evaluation design addresses matched resource budgets, retention of critical capabilities, external holdouts, adaptive selection, transfer, and recovery. Its central methodological requirements include separating runtime and development budgets and evaluating an accepted version on evidence that was not used to select it, unless a suitable selection-aware method is specified. The contributions are operational distinctions, continuity constraints, and a proposed empirical protocol. No new experiments are reported, and the framework does not establish open-ended evolution, consciousness, or artificial general intelligence. Author-released conceptual and methodological preprint in English. Not peer reviewed. No new experimental results. AI assistance, including English translation and language editing, is disclosed in the manuscript.
Authors
- Huajun Liu
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-10-29
- DOI
- https://doi.org/10.5281/zenodo.23752224
- Primary Topic
- Psychiatry, Mental Health, Neuroscience
- Type
- preprint