Naavya Shetty

Contemporary Artificial Intelligence (AI) systems often engage every input indiscriminately, resulting in unnecessary computation, unpredictable generalisation, and brittle behaviour on unfamiliar tasks. We present the Preprocessing Metacognitive System (PMS) 2.0, a system-agnostic metacognitive layer that evaluates incoming tasks and decides whether to accept, escalate, or refuse them before invoking any downstream reasoning system. PMS 2.0 seeks to provide interpretability at the level of computational governance - making transparent why the system chooses to engage, escalate, or refuse a task utilising confidence, task complexity, feasibility, novelty, and predicted benefit of escalation - to guide principled routing decisions without modifying downstream models. The system preserves conditional accuracy on escalated inputs, reduces computational load, and operationalises abstention as a first-class outcome, with previously refused tasks contributing to experience-informed efficiency gains. Evaluated across multiple domains and downstream architectures, PMS 2.0 demonstrates that metacognitive preprocessing can improve computational efficiency, reliability, and transparency, providing a practical framework for allocating resources where deliberative computation is most justified.

Machine cognition is currently heavily speed-based. Directly tackling inputs with computation often leads to inefficient steps, such as performing redundant or repetitive computation, or execution without assessing whether a task is within computational capacity. This paper proposes a preprocessing metacognitive system to be implemented in a manner such that it screens all input requests, creating a strategic ‘bottleneck’ to filter, redirect or halt the flow of control before computation begins. The findings theorise improved accuracy, reliability and resource-management, strengthening the argument for making metacognition an essential component of artificial intelligence.