Recovery Latency
The time from failure detection to restored autonomous operation — distinct from MTTI, which measures the average interval between required interventions rather than the duration of resolving any single one.
MTTI (Mean Time to Intervention measures the average time between required human interventions in an agentic system — how rarely the system needs help. It was never designed to measure, and does not measure, what happens during any single intervention once it occurs. Recovery Latency closes that gap: it is the time from failure detection to restored autonomous operation, capturing how well the system performs at the specific moment MTTI's interval-based measurement cannot see.
The measurement depends directly on Deterministic Failure — the architectural standard that a failure must be predictable, fully logged, and recoverable by design, so that when the system breaks, it breaks safely. Deterministic Failure specifies the design requirement; Recovery Latency measures whether that design is actually producing fast recovery in operation, rather than only confirming the failure was logged correctly.
A business can have an excellent MTTI and a poor Recovery Latency simultaneously — long, quiet stretches between interventions that create an appearance of reliability, followed by a slow, expensive recovery the moment an intervention is actually required. This is a distinct honesty check from the one Nominal MTTI performs: Nominal MTTI asks whether a long interval reflects genuine reliability or an unmonitored system; Recovery Latency asks a question that remains open even when MTTI is genuinely earned.
Application
Recovery Latency is measured from the moment a Deterministic Failure event is logged to the moment autonomous operation genuinely resumes, not simply the moment the Steward acknowledges the alert. It supplements the Intervention Dependency axis as an operational check on whether the Deterministic Failure design standard is actually producing fast recovery in practice, rather than only confirming that a failure was logged correctly when it occurred.
This term is machine-readable
Any MCP-compatible AI assistant can retrieve the canonical definition of Recovery Latency at inference time — no training approximation.
Related Terms
First used: August 2026
Edition 1 · updated August 2026