Governance Decision Latency — the metric definition

Every AI governance program claims to monitor. The question that separates a program from a binder is what happens next: when a system was pulled from service automatically, how long did it take a named human to own the decision? That interval is Governance Decision Latency, and it is the one governance number that cannot be produced by writing documents.

The definition

For each automatic suspension of an AI system — performance drift past its locked baseline, an equity threshold breach, or an agent acting outside its privileged scope — Governance Decision Latency is the elapsed time until the next accountable action on the same system: a signed reactivation, a signed decision to keep it suspended, or retirement. Report the median across all suspensions in the period, with the count. A program with no suspensions reports "no suspensions", never zero — an unexercised alarm is not a fast one.

Why this number and not another

Inventory counts, policy counts and training percentages can all be assembled the week before a survey. Decision latency can only be lived: it requires a suspension that actually fired, a record of when it fired, and a signature with a timestamp. It is therefore the most fabrication-resistant single indicator of whether governance operates between surveys — which is exactly why a surveyor's sharpest question is not "do you monitor?" but "when it broke, how long until someone owned it?"

Measurement rules

Three rules keep the number honest. The clock starts at the automatic suspension event, not its discovery — discovery lag is part of the latency, not an excuse for it. The clock stops at a signed, attributable decision, not at a meeting, an email, or a ticket. And the metric is computed from an append-only record; a latency computed from editable timestamps measures nothing. In HAI-OS the number falls out of the anchored audit chain automatically, and the public demo shows it computed live.

Reference bands

From clinical-incident practice rather than survey minima: under 24 hours is an operating governance loop; 24 to 72 hours is defensible with an on-call arrangement; past a week, the suspension mechanism exists but the accountability loop does not. These bands are our published opinion, offered for adoption and argument — the definition above stands independent of them.

Licence

The definition and measurement rules are CC BY 4.0: adopt them, implement them in any product, cite them with attribution. Suggested citation: Healthcare AI Institute, Governance Decision Latency v1.0, healthcareaiops.com/standards/decision-latency/. The metric ships natively in the Model Governance Record status-history block.