# Measurement is not authority

> Measurement is not authority means a computed score, rank, or state is a measurement that informs a decision — it is not permission to act. A named human holds the authority to decide.

Category: Governance
Also searched as: metric vs mandate, human-in-the-loop, decision rights
Source: Mediator Solutions — https://mediatorsolutions.io/learn/#measurement-is-not-authority
License: free to read, learn, cite, and apply, with attribution to Mediator Solutions.

## What it is

As systems compute more state, it becomes tempting to let the computation decide. That collapses the line between measuring a situation and being authorized to act on it. The discipline keeps them separate: the engine computes and holds state; a model may advise; a named human decides. A rank is a measurement, not an authority.

## Why it matters

As systems compute more, the temptation to let the computation decide grows, because it is faster and feels objective. That is precisely where accountability disappears: when a number acts, no one is answerable for the action. The discipline keeps a named human in the loop not as ceremony but as the locus of responsibility — someone who can be asked why. The subtle failure is authority creep: a measurement that was advisory last quarter quietly becomes the decision this quarter because no one re-examined the boundary.

## When to use it

- A computed score, rank, or state is about to trigger an action automatically.
- A model’s recommendation is being treated as a decision rather than advice.
- An advisory metric from last quarter is quietly driving decisions this quarter.

## Principles

- A model advises. An engine computes state and a receipt. A named human decides.
- A rank, score, or state is a measurement, not permission to act.
- The authority to act is held by a person who can be named, not by a number.
- Fail-closed: when authority is unclear, the system holds rather than acting.

## Practice

1. For each automated output, ask: is this measuring, or is this deciding?
2. If it is deciding, route the binding act to a named, authorized human.
3. Keep the measurement, the recommendation, and the authorization as separate records.
4. Default to hold when the authority for an action is not explicit.

## Where it fails

- **Authority creep** — A measurement that was advisory becomes the decision over time because no one re-examines the boundary, and accountability silently disappears.
- **Accountability vacuum** — When a number acts, no human is answerable for the action, so a wrong outcome has no owner.
- **Objectivity illusion** — The computation feels objective, so its output is trusted past the point where a human should have weighed context it could not see.

## In practice

An engine ranks counterparties by risk, and the top-ranked one is about to be cut off automatically. The discipline inserts a named human: the rank is a measurement that routes the case to a decision-maker, not a decision itself. The human can see what the rank cannot — a known data gap, a relationship, a pending correction — and is answerable for the cut-off either way. The engine computes and holds state; the human decides and owns it.

## Verification

For any consequential action, the record shows a measurement, a recommendation, and a separate authorization by a named human — three distinct artifacts, not one blended step.

---

Previous: https://mediatorsolutions.io/learn/#position-before-action
Next: https://mediatorsolutions.io/learn/#decision-hygiene