Who asked the model — a request trail you can actually look at

TL;DR. If an organisation lets people call a model, someone will ask: who called what? A useful answer is an operator request trail — filter by team, route, and status — without dumping the contents of the prompt. That is not a full compliance archive. It is not Eval. It also does not replace sandboxes around tools an agent may use.

The question that arrives after the first “yes”

A bureau or a large team often starts with a generous sentence:

“Staff may use the AI — here is the endpoint.”

A week later, a quieter question shows up from IT, security, or the person who will sit with an auditor:

“Can we see who used it?”

That is not the same as “show me every prompt.” Dumping contents is a privacy problem of its own. What operators usually need first is a traffic log they can actually open: who, which route, did it succeed, when.

Think of a building’s visitor book, not a tape of every conversation in every room.

What the trail is (and is not)

Need	Request trail (this)	Something else
Who called which model route?	Yes — user / team / key, route, status	—
Filter the noise	Team, route, status class (success vs client / server error)	—
Read the prompt and the reply	No — the operator listing does not return payloads	A heavier records policy, if you ever need that
Keep years of evidence for a regulator	No — this listing is recent history (hours to a week), not a forever archive	Records / legal hold is a separate design
Prove the answer quality before a change ships	No	OrchAI Eval — the quality gate
Stop an agent leaving the test box	No	Sandboxes and human gates (a different literacy topic)

The honesty line matters. “We have audit” is a slide. “An operator can filter who called which route, without opening the prompt” is a product sentence.

What an operator can actually do

On OrchAI Portal, the model gateway has a Request logs page for this job. An operator can:

• See who called which route, and whether it came back success, a client error, or a server error.
• Filter by team, route, and status.
• Stay inside the keys they are allowed to read — not a dump of every credential in the building.

They cannot, from that listing, scroll a transcript of what was typed. That is the point: visibility without vacuuming the contents.

The listing is also short-horizon. It is meant for “what happened recently,” not “export seven years for the file room.” If a buyer needs a true records archive, that is a different conversation — do not treat this page as that archive.

Two different “can we govern this?” questions

People often smash three fears into one word, audit:

1. Who used the front door? — request history.
2. Is the answer still good? — eval / test suites.
3. What may the agent do with tools? — sandboxes, permissions, human gates.

This post is (1). (2) is Eval — a quality gate, not a log viewer. (3) is the next literacy beat after a model alone is not an agent: when tools exist, “try harder” can mean leave the test box. A request trail complements that; it does not replace containment.

How this connects to what we build

OrchAI Portal is the enterprise runtime: one login, roles, and a governed place to work — including the model gateway that teams actually call. Request logs are operator chrome on that gateway, not a new product and not a fifth pillar. Library still holds cited knowledge; Agents is still the front door; Eval still watches quality.

If your organisation has already said “people may call the model,” and the next meeting is “can we look at who did,” talk to us or try the demo — and bring the visitor-book question, not a request to paste every prompt into a slide.

Frequently asked questions

Is this a full compliance archive? No. It is recent operator history, filterable, without a prompt dump. Forever records are a separate design.

Is this Eval? No. Eval tests whether quality still holds before a change ships. Logs say who called the gateway.

What can an operator see? Who / team / key, which route, status, timing and usage metadata — not the prompt or the reply.