Na minha lista:
| Autor principal: | |
|---|---|
| Formato: | Recurso digital |
| Idioma: | inglês |
| Publicado em: |
Zenodo
2026
|
| Assuntos: | |
| Acesso em linha: | https://doi.org/10.5281/zenodo.19600943 |
| Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Sumário:
- <p>The Persistence Kernel is a minimal, executable control framework for enforcing alignment in autonomous, multi-agent, and distributed systems through structural constraint rather than optimization or behavioral training.</p> <p>Unlike conventional approaches to AI alignment—which attempt to shape or optimize model behavior—the Persistence Kernel operates as a runtime substrate that governs whether system transitions are admissible. It enforces alignment as a condition of continued existence: only trajectories that remain within a defined region of viability are permitted to proceed.</p> <p>At the core of the framework is the concept of a <strong>viability window</strong>: a bounded region of state space within which the system can persist under constraint. At each step, the kernel evaluates candidate transitions proposed by one or more agents (including large language models) against a governing rule (“law”) informed by the system’s current position relative to the boundary of this window.</p> <p>A structural signal—derived from measures such as boundary distance, distortion, and constraint pressure—modulates the strictness of admissibility:</p> <ul> <li>Far from the boundary, the system admits a broader range of transitions.</li> <li>As the boundary is approached, constraints tighten.</li> <li>Near the edge, only stabilizing transitions that preserve or restore viability are permitted.</li> </ul> <p>This enables <strong>continuous self-regulation</strong>: the system compensates as it approaches failure, favoring actions that maintain structural integrity without requiring explicit optimization, reward functions, or goal-directed behavior.</p> <p>When no admissible continuation remains, the kernel invokes a <strong>collapse operation</strong>. Collapse is not failure or termination, but a deterministic transition to a reduced, safe regime that preserves system identity while restoring definability and admissible continuation where possible. If no such continuation exists, the system halts in a well-defined terminal state.</p> <p>Key properties of the Persistence Kernel include:</p> <ul> <li><strong>Eliminative constraint enforcement</strong> (no ranking, scoring, or optimization)</li> <li><strong>Append-only, irreversible history</strong> ensuring full auditability and replay</li> <li><strong>Identity preservation</strong> across all admissible transitions</li> <li><strong>Boundary-aware governance</strong> via viability-informed law modulation</li> <li><strong>Deterministic collapse semantics</strong> for safe degradation under constraint exhaustion</li> <li><strong>Multi-agent compatibility</strong> without reliance on consensus, voting, or shared optimization</li> <li><strong>Distributed operation</strong> with local authority preserved across nodes</li> </ul> <p>The framework is implementation-agnostic and can be applied to any system in which state transitions must remain within constrained bounds, including AI agents with real-world actuation (e.g., filesystem access, APIs), infrastructure automation, financial systems, robotics, and long-running adaptive processes.</p> <p>By shifting alignment from model behavior to system-level admissibility, the Persistence Kernel provides a structural approach to safety: it ensures that unsafe trajectories are not merely discouraged, but <strong>non-continuable</strong>.</p> <p>This work presents both the formal specification and an executable reference implementation of the Persistence Kernel, demonstrating that alignment can be enforced as a property of constrained persistence rather than learned intent.</p>