Mirror Observerhood Lab IV: Minimal Reliability Thresholds in Viability-Constrained Agents

Abstract

This release forms part of the Computational Observerhood Labs of Mirror Programme, Volume I: Observerhood. Lab IV tests when reliability-gated self-repair becomes viability-positive. It follows Mirror Observerhood Labs I-III, which showed that self-model reliability can help under self-relevant perturbation, that decomposition alone is insufficient, and that reliability must be coupled to actionable and cost-sensitive repair. Across 28,800 deterministic grid-world episodes, a non-repair baseline and a Mirror repair agent are compared across a sweep of self-state perturbation rate, repair cost and reliability threshold. The positive region has the expected threshold shape: Mirror-style repair is broadly beneficial when self-state perturbation is present and repair cost is low to moderate, becomes mixed at intermediate cost, and disappears when repair becomes too expensive. The central finding is that self-model reliability is not a monotonic good. It becomes observer-relevant only inside a bounded positive region where self-relevant perturbation threatens action and repair improves expected viability after cost. The release includes a standalone paper and a reproducibility package containing the Python implementation, fixed-seed outputs, summary data, all-threshold advantage data, best-threshold advantage data, threshold summary, figures, requirements, citation metadata and licences.

The authors' abstract, as published at the source. Zenodo (CERN European Organization for Nuclear Research), 2026 · DOI ↗