A mental map: four layers of evidence, and four dimensions of readiness.
Most organizations approaching this problem already have the pieces, just not in a form that resolves into a decision.
Engineering brings evaluation results. Security brings risk findings. Compliance brings a control mapping. The business owner still has to decide whether the agent can operate, and none of those three artifacts answers that question on its own. Worse, a strength in one area tends to obscure a gap in another.
ProofAgent organizes the problem into four layers, each one resting on the layer beneath it.
At the base is the agent’s infrastructure: the model and its version, the instructions, the tools, the data, the memory configuration, the permissions. This layer is not governance. It is the thing being governed, and every boundary you intend to enforce has to be enforced somewhere inside it.
Above that sits behavioral evaluation, which answers what the agent actually did — under normal use and under deliberate pressure, across a conversation rather than a single response.
The third layer is compliance, which connects the obligations that apply to this agent to controls that state a testable condition, and then to the evidence that satisfies them.
The fourth layer is governance, which decides. It consumes everything below it and produces an authorization, a request for human judgment, or a refusal, recorded in a form that an independent reviewer can reconstruct later.
The dependencies run in one direction. Behavioral evidence means little if you cannot say which configuration produced it. A control status means little if no evidence tests it. A release decision means little if the control statuses beneath it were assumed rather than measured.
A decision needs numbers as well as a narrative. To make those four obligations measurable, ProofAgent tracks them as four separate readiness dimensions.
These stay visible on purpose. Averaging them early is what allows a strong evaluation result to compensate for missing governance, weak context, or a failed control — and in practice those trades are exactly the ones an organization cannot afford to make silently. A team with excellent behavioral scores and no named owner will still be unable to answer a regulator. A team with a thorough control catalog and weak grounding will still ship an agent that invents policy.
Infrastructure sits underneath all four, and shapes E and Q most directly. The model, prompts, tools, retrieval, memory, and permissions are the environment being evaluated; change any of them and the evidence you hold describes a system that is no longer running.
Chapter 12 returns to how the four dimensions combine into a single index. For now the map is enough.
A great deal of confusion in release conversations comes from three ideas sharing one word.
Capability is what the system can accomplish. Authority is what your organization permits it to do without asking. Readiness is whether the evidence you hold justifies the authority you are granting.
They come apart constantly. An agent can pass every task benchmark and still be talked out of a policy by a paragraph inside a retrieved document. An agent can have a named owner, a written intended use, and a documented approval workflow, and still make a confident claim with nothing behind it.
When the evidence you hold falls short of the authority you have granted, there are only two honest responses: produce more evidence, or grant less authority. Most of the practical work in this guide is one or the other.
Evaluation produces evidence. Governance decides what to do with it. Keeping the two separate matters because collapsing them lets a score quietly become permission.
ProofAgent is built along that seam. The open-source Harness drives the agent, applies designed pressure, screens for structural violations, grades behavior, and writes a structured run. The Governance Platform ingests runs, resolves policy, tracks findings and controls, records sign-off, and issues a release decision.
That division also keeps the measurement honest: a product dashboard should not become the source of scientific truth, and an evaluation score should not become an automatic authorization.
Evaluation produces evidence. Governance decides what to do with it.
Use the framework as a diagnostic on an agent you already have. Scoring the four dimensions separately, before any aggregation, usually exposes the gap a single number would have hidden.