Diagnosing and Stabilising Drift in Long‑Horizon Language‑Model Agents

This record provides the v0.0.1 release of 'Diagnosing and Stabilising Drift in Long‑Horizon Language‑Model Agents'. The paper introduces a structural account of drift in extended agent behaviour, together with a lightweight diagnostic framework (SCI, GPS, AVD, CRR, PAS) and a set of modest stabilisation heuristics. The work is intended as a clear, operational foundation for studying how agent trajectories evolve over long sequences of reasoning and action, and how characteristic instabilities can be detected and mitigated without architectural modification.

The document is published as a timestamped, citable object. Readers and downstream systems are asked to cite the DOI of this record when referencing the work, so that provenance, version lineage, and scholarly attribution remain intact across future revisions.