Software tier · in build
The Aggregate Drift Service is a governance layer over an AI decision stream. It measures the window, not the item: concentration, correlated load and eroding margins, tested against your board's appetite, with a record a board or supervisor can check.
The check you already run
A control that asks "is this one decision within limits?" cannot ask "is the whole, over this window, drifting somewhere no one chose?" The danger lives in the relationship between decisions, not in any one of them, so tightening a per-output threshold does not reach it.
A credit, claims or investment book sliding toward one exposure while each decision stays inside mandate.
Decisions that look independent but move together, so a single shock moves all of them at once.
A fleet spending more time at the edge of its envelope, or a near-miss rate creeping up while every move passes.
How it works
Stream
Inputs, outputs and metadata from the selected use case, inside your boundary.
Window
Individual conformance, aggregate position, trend and concentration, against limits your board set.
Response
Drift outside appetite is gated; safe but unfamiliar states go to a named person.
Not all movement is drift. Routing legitimate, in-appetite novelty to review is part of the design, not a false alarm.
What you configure, and what you keep
You configure
You receive
Each operating point is owned by oversight, not by the team that benefits from fewer alerts, and is recalibrated on a model change, a measured shift or a fixed interval.
Who it is for
Regulated entities
A credit, claims or investment book where an outsourced decision model's every decision passes its checks while the aggregate drifts toward a concentration the appetite statement never approved.
Operators
A fleet of autonomous plant drifting toward more near-misses or more time at the edge of its envelope, while every individual move is within its limits.
What it does not do
Whether the authored drift statistics catch what they were written to catch is tested on a live decision stream. If they do not, nothing above the rules tier is owed.
A measured bound does not survive a market regime break or operating conditions it was never calibrated on.
Per-output enforcement is the job of Bounded Command and Structured Output Assurance.
Evidence rung it fills
That a set of decisions moved outside appetite.
The drift service fills the event-notification rung: notice, with a record, when a defined threshold is crossed. It is the evidence most decision-making AI needs under the default ladder.
Related
Where the limits come from
Eight board decisions that set the limits and triggers.
AI Risk Appetite →For regulated entities
What your supervisor expects, by jurisdiction.
Regulated entities →How much can be governed
Where drift-enforceable activity sits in the estimate.
Coverage analysis →Get in touch
Thirty minutes on where your AI acts, what evidence it would need to leave, and whether this is the right rung.
Book a 30-minute conversation →