Inference as Training

Inference is usually described as fixed-model execution: an input is processed, an output is produced, and the system that performed the inference remains unchanged. That description is useful for ordinary feed-forward evaluation, but it is incomplete for biological, adaptive, and agentic systems. Some inference episodes do not merely produce an answer. They resolve mismatch by changing context, routing, temporary state, memory, gain, synaptic efficacy, or plasticity markers in ways that affect later inference. This paper defines "inference as training" operationally. An inference episode is training-like only if resolving mismatch at time `t` changes a measurable system variable that affects inference at `t + k`. The criterion is residue, not rhetoric. A system must leave a measurable after-effect, and that after-effect must causally alter a later inference episode under appropriate controls. The paper rejects three stronger claims: that all inference is training, that biological inference is standard artificial-neural-network backpropagation, and that AI test-time training validates a neuroscience theory. Instead, it proposes a shared measurement language for biological plasticity, predictive feedback, agent memory, test-time adaptation, and self-correcting inference. Thermodynamic dissipation remains a modeling analogy for local mismatch reduction. Biological feedback is treated as feedback-driven plasticity. AI systems are compared by what changes during and after inference, how long that change persists, and whether disabling the residue removes the later benefit.