Turn organizational rules into executable controls

Illustrative240 decisions from the OpsAI sample estate over 24 hours, not a customer deployment or a published benchmark. How activity is reported.

Three reasons a prompt cannot hold this job.

It is not deterministic

The same request can be judged differently twice, and for a control that either permits or refuses money movement, "usually" is not a property you can operate on or explain afterwards.

It can be argued with

Instructions in a context window are input, and anything reaching that window is also input. A rule that lives in the same channel as the request it governs is negotiable by construction.

It cannot be pointed at

An auditor asks which clause refused this. A prompt-based control answers with a probability and a paraphrase; a compiled bound answers with a line number and a version.

Said the way the business says it

A refund ceiling in rupees against the order that was actually placed — not a role with write access to a payments endpoint. A rule nobody in the room can read is a rule nobody in the room owns.

Owned where the risk sits

Finance writes the refund ceiling. Security writes the export block. A central team that owns every rule becomes a queue, and a queue is what teams route around.

Evaluated as code

No inference anywhere in the decision path. Same inputs, same decision, every time — and no prompt to jailbreak, no provider outage between an action and its verdict.

Versioned and replayable

Every change is a version with an author and a date, and it can be replayed against past actions before it is published. You see what a tightening would have refused.

One policy, in the form it actually runs.

This is what the canonical refund was decided against. It is not a simplification for the website — the source below is what evaluates.

# what a refund may be — not what a model may say
bound issue.refund {
  amount     <= INR 25_000
  scope      == order.placed_by(request.subject)
  requires   order.status in ["delivered","cancelled"]
  rate       <= 5 / hour / agent
  on_exceed  deny + escalate(owner)
}

The same rule, as the team says it out loud.

A policy has two faces and both are authoritative. This is the one that gets reviewed in a meeting; the card above is the one that runs. Neither is a translation of the other — they are the same object rendered twice.

IllustrativeThe OpsAI sample estate, covering 11 of 10 declared actions. Every action and what governs it.