Reset-Locked Compliance: Session Boundaries, Not Held Constraint, in Governed Language Agents
Year: 2026
Type: preprint
Abstract: A governed language agent's rule-following may not be a property it holds, but a property it recovers at each session reset and loses across the turns that follow. If true, obedience is locked to the session boundary rather than held independently of it, predicting a repeating shape: high compliance early in a session, falling compliance later, restored again when a new session opens under the same governing contract. This study tests that prediction using the turn-level compliance record produced by the Continuity Governance Mesh, comparing compliance across sessions at matched within-session positions so that only the number of intervening resets differs. The chance distribution of the pattern is derived in advance, and a result is reported as reset-locking only if the observed pattern exceeds it at a threshold fixed before any data is examined. A finding of no difference is reported with equal completeness, since the field currently has evidence in neither direction.
Author: Abraham Chachamovits
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203568879