Ask what makes a system trustworthy and the answers arrive as adjectives: hardened, compliant, audited, aligned. Ask instead what would let you trust a specific system — this afternoon, with your name on the approval — and the answer gets concrete fast. You want to see what it is doing. Which controls are on. Where its output came from. What it declined to do, and why.
At the point of use, trust is an interface problem.
Governed and legible are different properties
Most production systems carry far more governance than they show. Somewhere behind the response there is a policy engine, an approval gate, a provenance trail in a log aggregator. At the moment a person has to act on the system’s output, almost none of that is visible. The operator approving an agent’s change sees the change. They do not see which policies evaluated it, under what identity it ran, or which of its inputs came from a source anyone vetted.
People fill that gap with posture. The security page lists certifications the way a diner frames its first dollar bill — sincere, and no help at all with the decision in front of you.
The companion note defines trust operationally: policy plus evidence at the boundaries where control changes hands. This note adds the corollary: policy and evidence only function as trust where a person can see them. A control that fires invisibly protects the system. A control that fires visibly lets someone rely on it.
What a trust surface shows
Three kinds of structure belong at the point of use, and every one of them is data the system already has.
Which controls applied. A policy decision has an identity: which policy, at which version, evaluated this request, and what it concluded. Decision logs in engines like Open Policy Agent already capture exactly this shape — the work is carrying it forward to the surface where the decision lands, instead of leaving it in an archive.
Where the output came from. Provenance has a standard vocabulary — W3C PROV models exactly the who-derived-what-from-what chain — and the consumer-facing version already exists in C2PA content credentials , which travel with the artifact rather than in a log beside it. That distinction is the whole point. Provenance that ships with the output can be checked at the moment of use; provenance that lives in a warehouse can be checked in a postmortem.
What the system declined. Refusals are the strongest trust signal a surface can carry, because a visible refusal is the control demonstrating itself. A system that shows only what it did invites the question of what it would have done. A system that shows “two of seven retrieved sources were excluded by policy X” answers it.
Visibility is a design constraint, not a dashboard
The difference between an audit log and an interface is who it serves and when. Logs serve reconstruction after something went wrong. Surfaces serve reasoning while the decision is still open. Building for the second turns out to impose useful discipline on everything upstream: a policy you intend to display needs a stable name and a machine-readable verdict; provenance you intend to display has to survive every handoff between the source and the screen, which is how gaps in the chain get found. A policy that cannot be displayed is usually a policy that cannot be audited either — display just surfaces the debt earlier and cheaper.
Telemetry infrastructure makes this less work than it sounds: OpenTelemetry already propagates context across service boundaries, and a policy identity or provenance reference rides in a span attribute as easily as a latency number does.
Where to start
Pick the one boundary your users already worry about — for most teams right now, that is the point where an agent’s output enters a production system — and put two things on its surface: the identity of the policy that approved the action, and the provenance of the inputs it acted on. One boundary, made legible, teaches you more about your evidence gaps than a quarter of dashboard work.
The Agent Audit Packet is the structured version of this exercise: a boundary-by-boundary worksheet for finding where your controls are real but invisible.