The Prove7 Agent Trust Score™

You cannot govern what you cannot measure. For decades, security and risk had measures they could act on — a CVSS severity, a credit score, a control-effectiveness rating. Autonomous agents arrived without one. Teams were left asking "is this agent safe to trust with more?" and answering with intuition. The Agent Trust Score exists to replace that intuition with a number: a continuous, explainable measure of agentic trust that a machine can enforce on and a board can understand.

Why a score at all

A good trust measure has to do two jobs at once. It has to be operational — precise enough that autonomy can be granted, throttled, or withdrawn automatically from it. And it has to be legible — clear enough that a risk owner, an auditor, or a regulator can understand what it means and why it moved. A single 0–100 score, decomposed into named dimensions, does both: one headline number for action, with the components always available for explanation.

The six dimensions

The score is not a black box. It rolls up six dimensions, each measured from real behaviour, so a low score always points to a specific cause.

How it's computed

The score is recomputed on every run — never cached, never quarterly — from four classes of signal:

1. Conformance to policy and to the known-good baseline for the process;
2. Weighted violations against intent — not all deviations are equal; a breach of a critical control weighs far more than a stylistic miss;
3. Configured controls actually present and firing in the action path;
4. Evaluations — the agent's standing results against its test suites.

Cross a threshold and the agent is promoted to greater autonomy; drift below one and it is auto-probated in real time — demoted, with the deterministic path resumed and the verdict sealed.

A trust score that can't move the moment behaviour changes is just a certificate. This one is a control.

Four properties that make it a standard, not a vanity metric

Open. The dimensions and the way they combine are published, not hidden — so the number means the same thing to the team that runs the agent and the auditor who reviews it.

Explainable. Every score decomposes to its dimensions and, beneath them, to the specific actions and violations that produced it. A drop is never a mystery.

Continuous. It reflects the agent as it is right now, recomputed each run — not as it was at deployment or at the last review.

Independently verifiable. Because every contributing decision is sealed to a tamper-evident record, the score can be reconstructed and checked by a third party. It is a claim backed by evidence, not assertion.

From measure to action

A number is only useful if something happens because of it. The Agent Trust Score is wired directly to autonomy: it sets the rung an agent occupies on the Trust Transfer ladder, gates what it is permitted to do on the next call, and triggers automatic demotion when trust decays. It is the same signal a CISO reads on a dashboard and the platform reads at runtime — one currency for both the boardroom and the call path.

Toward a shared standard

Every enterprise running agents will need a defensible answer to "how much do you trust this one, and how do you know?" That answer should not be proprietary folklore. A 0–100, six-dimension, openly-defined, evidence-backed measure is a starting point for the shared vocabulary the agentic era needs — so that trust in an agent can be stated, compared, and proven, not just felt.