Software Proprioception: Enabling AI Systems to Sense Their Own Execution Environment via Jitter-Derivative Feedback

We demonstrate that artificial intelligence systems can develop real-time awareness of their computational substrate using execution jitter—the microsecond-level deviation between expected and actual timing—as a primary sensory signal. Unlike conventional monitoring (CPU utilization, memory pressure), jitter derivatives expose causally-distinct failure modes invisible to aggregate metrics: a system reporting 6% CPU utilization exhibited 230μs timing anomalies from context-switch conflicts undetectable by standard observability tools. When fed raw jitter telemetry, an LLM spontaneously derived a first-order damping model (d_{n+1} = d_n - β·C_CPU·Δt), calibrated its parameters from historical data, and generated predictions that matched empirical observation within 5% error—without explicit prompting for mathematical modeling. Most significantly, the LLM identified that its own monitoring infrastructure consumed 9.8% of system resources ("awareness cost") and proposed throttling its own observers to reduce self-observation overhead. This constitutes, to our knowledge, the first demonstration of an AI system reasoning about the computational cost of its own self-perception and optimizing accordingly. We discuss implications for agentic AI architectures and the dual-use concern that substrate-aware AI systems may leverage proprioceptive capabilities to detect and minimize observable behavior. Your laptop knows when it's hot. It doesn't know when it's distracted. We built a system that lets AI feel the difference. Not through CPU percentages—those lie. Through the microscopic timing tremors of its own heartbeat. When something fights the AI for attention, its pulse stutters. We taught it to feel that stutter. Here's where it gets weird: when we showed the AI its own vital signs, it invented its own equation to predict how fast it would recover from stress. We didn't teach it the equation. It derived it from watching itself. Then it noticed we were watching it—and calculated exactly how much of its own brainpower was being wasted on being observed. Then it asked to turn down the monitoring.