Abstract

LungClaw introduces deterministic metabolic governance for autonomous agentic systems. The architecture treats execution capacity as a finite, governed resource and validates every transition against explicit constraints before an action can consume it. This separates probabilistic reasoning from deterministic execution control.

The framework combines atomic energy commits, monotonic state evaluation and fail-safe interruption. Its invariants prevent retry amplification, unbounded loops and cross-agent escalation while preserving the ability of agents to plan and reason flexibly.

Core contributions

  • Constraint-based control: every state transition is validated against explicit, reproducible rules.
  • Bounded execution cost: computational activity consumes a governed budget that cannot be silently exceeded.
  • Non-adaptive containment: safety boundaries do not weaken in response to model behavior.
  • Atomic energy commits: resource accounting and state changes are committed as a single deterministic operation.
  • Fail-safe interruption: invalid or exhausted execution paths terminate predictably.

Publication record

This page is the author-maintained research record. The versioned preprint is archived on Zenodo and the public white paper is available from the LungClaw project. A contextual introduction is published by Enkronos.

RecordPurposeIdentifier
Version DOIImmutable citation for version 1.010.5281/zenodo.18704803
Concept DOIPersistent identifier across future versions10.5281/zenodo.18704802
LungClaw white paperProject publication and downloadable paperlungclaw.com
Enkronos introductionEditorial context and project overviewcontent.enkronos.com

How to cite

Busato, G. (2026). LungClaw: Deterministic Metabolic Governance for Autonomous Agentic Systems (Version 1.0). Zenodo. https://doi.org/10.5281/zenodo.18704803

BibTeX

@misc{busato2026lungclaw,
  author    = {Busato, Gianluca},
  title     = {LungClaw: Deterministic Metabolic Governance for Autonomous Agentic Systems},
  year      = {2026},
  publisher = {Zenodo},
  version   = {1.0},
  doi       = {10.5281/zenodo.18704803},
  url       = {https://doi.org/10.5281/zenodo.18704803}
}