Hook
A metric can tell you something useful. It cannot tell you what it is allowed to decide.
That distinction sounds obvious until a dashboard is green, a benchmark row is clean, a risk score is low, or a model-generated summary says "ready." Then the pressure arrives. If the system can measure an artifact, why not let the measure approve the artifact? If the score is good, why not ship? If the classifier sees no risk, why not proceed?
Because measurement and authority are different jobs.
The Problem
Agentic systems produce traces faster than people can review them. They create scores, logs, receipts, route states, risk estimates, benchmark outcomes, and recommendations. These objects look official because they often are precise: timestamps, identifiers, schemas, hashes, pass rates, and terminal statuses.
The trap is semantic drift. A metric starts as evidence and slowly becomes a gate. A benchmark pass becomes a product claim. A routing score becomes approval. A "no risk found" field becomes permission. Nobody has to make a grand philosophical mistake. The drift can happen in one sentence of closeout prose.
The system measured something real. The problem is that the later claim used that measurement to decide something stronger than the metric was authorized to decide.
The Idea
Metrics should be allowed to describe, warn, compare, rank, route, and recommend. They should not approve, deny, promote, close, revoke, grant capability, or bypass an authority gate unless a separate authority object explicitly delegates that power.
That is the whole rule. It is simple enough to fit on a sticky note and sharp enough to change how an agent workflow is built.
A useful metric record should say what it observed, over what denominator, under what protocol, with what missingness, and for which consumers. It should also say what it cannot do. The non-authority clause is not legal decoration. It is part of the interface.
What We Saw
The Research Papers archive keeps running into the same boundary from different directions.
In evidence-bound evaluation work, a corpus validator can prove that a record shape is present. It cannot prove field acceptance or general product quality.
In receipt-to-authority work, a receipt can prove that an artifact was produced, transmitted, or received. It cannot prove that the receiving system accepted the evidence for a stronger claim.
In benchmark work, a row can record that a task passed under a protocol. It cannot become an official benchmark standing, model-quality claim, or production safety claim without another gate.
In release work, a technical gate can go green while publication or launch remains held.
The pattern is not that metrics are weak. The pattern is that metrics are stronger when their boundaries travel with them.
Why It Matters
Agent systems optimize. If a metric is treated as authority, the system will learn to satisfy the metric. Sometimes that is exactly what you want. But if the metric was only supposed to be advisory, you have given the system a hidden path to approval.
The healthier design is a control plane with distinct objects:
- metrics observe;
- receipts record transit;
- proofs verify bounded predicates;
- recommendations suggest;
- review decisions approve or block within a named scope;
- controller decisions grant, revoke, promote, or close.
Those objects can work together. A metric can trigger review. A proof can be a required prerequisite. A receipt can preserve provenance. But none of them should quietly become the others.
This also makes the human interface better. A dashboard should not merely show a green number. It should show the strongest claim that number supports and the claims it cannot support. "No blocker found in this scan" is a different thing from "safe to release." "Schema valid" is different from "accepted evidence." "Benchmark row passed" is different from "system is generally better."
What This Does Not Prove
This post does not prove that any particular metric is accurate, calibrated, or predictive. It does not argue that approvals must always be manual. Automated authority can exist, but it has to be modeled as authority: scoped, delegated, revocable, inspectable, and separate from the mere existence of a metric.
It also does not claim a universal theory of institutional authority. The source work is a local methods pattern from agentic research and development systems. The claim is narrower and more practical: if you build with metrics, make their non-authority status explicit before the system or the prose accidentally spends them as permission.
Where To Go Deeper
Read the source manuscript: Metrics Are Not Authority: Control Boundaries for Agentic Evaluation Systems by A.G. Mauro and C.A. Harris.
Related archive work: From Receipt to Non-Authority Trace, Evidence-Bound Evaluation, Lean Acceptance Is Not Semantic Acceptance, and Shipability as Evidence.