AI Needs Instincts, Not Just Instructions: Allocating flexibility and guarantees in reliable AI systems

Language-model systems are increasingly capable but not correspondingly reliable. This essay argues that reliability depends less on making models deterministic than on deciding, requirement by requirement, which behaviors may vary and what actually keeps the rest fixed. Starting from a cautious analogy with biological control, where fast rigid responses coexist with flexible learning, it separates three often-conflated notions: determinism, behavioral disposition, and invariant. It argues that a requirement is an invariant only if something outside the model enforces it, and reviews evidence that guarantees placed inside an optimizing system are leaky. It surveys existing work on agent harnesses, behavioral contracts, compiled workflows and tool protocols, which address enforcement but not allocation, and proposes a five-question rubric for assigning each requirement to a learned disposition, flexible reasoning, an external invariant, or verification. The rubric is illustrated on a translation task and presented as a hypothesis with a test protocol. No new experiments are reported. This is a non-peer-reviewed preprint.