Abstract
Modern computational and artificial intelligence systems are typically evaluated based on constraint satisfaction at individual time steps. A system is considered correct if its current state satisfies a set of constraints or objectives. This paper argues that such state-based evaluation is insufficient. We show that systems may satisfy all locally observable constraints at every step while irreversibly losing the ability to return to correct states. These failures cannot be detected at any single time step. This introduces a distinction between feasibility (a property of states) and recoverability (a property of trajectories). We define a set of trajectory-level failure modes (including persistent deviation, recoverability failure, lock-in, and constraint observability failure) and demonstrate their relevance across modern systems, including AI pipelines, distributed systems, and software execution environments.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
pham, thom, "Trajectory-Level Failure Modes in Constraint Systems: Irrecoverability, Observability Failure, and the Limits of State-Based Monitoring", Technical Disclosure Commons, (April 09, 2026)
https://www.tdcommons.org/dpubs_series/9754