Detection and Decision Flow
Two related capabilities work together to keep every AI tool visible: detection finds tools nobody registered, and the decision engine decides what happens once a request arrives.
Default posture: allow, but never invisible
OpenProdkt is opt-in, not default-deny. Nothing is blocked automatically — every restriction comes from a policy you write:
- An unregistered asset is allowed, and the attempt is recorded as a discovery for your team to review.
- A registered asset with no matching policy is allowed.
- When multiple policies match, the most restrictive outcome wins:
denybeatsapproval_requiredbeatsallow.
This means nothing you haven't explicitly restricted gets silently blocked — but nothing runs invisibly either. Unregistered and ungoverned activity is always surfaced.
From detection to governance
- Detection observes traffic and surfaces tools nobody registered. It never blocks or approves anything by itself — it only queues the finding for a human to promote or skip.
- Promotion turns a discovery into a governed asset, ready to have policies attached.
- Policy evaluation runs for a registered asset that has at least one policy assigned, and produces the final outcome. A registered asset with no assigned policies is allowed with reason
no_policies_assigned, is recorded in decision history, and is not billed.
Why this matters
- Nothing enters your environment invisibly — unknown tools are surfaced, not ignored.
- You decide what to restrict. Governance is additive: you write policies for the workflows that need approval or denial.
- Promoting a discovery into an asset is the moment a tool becomes eligible for policy enforcement.
What gets recorded
| Outcome | Decision history | Audit event | Billed |
|---|---|---|---|
Unregistered asset (asset_not_registered) | No | Queued as a discovery | No |
Unknown asset_id (404 asset_not_found) | No | No | No |
No policies (no_policies_assigned) | Yes | decision.evaluated | No |
| No match, match, or approval required | Yes | decision.evaluated | Yes |
Policy evaluation error (503) | Yes (deny) | decision.evaluation_failed | Yes |
Validation error (400) | No | No | No |
A decision records authorization, not execution. When the caller sends a traceparent, the operation keeps the trace ID and the control plane's span ID for correlation; they do not change the decision and are not returned in history. See Audit for what is exported and how long it is kept, and Reason Codes for each reason.
Execution coverage is not a decision
After an allow, the control plane learns what happened only if the caller reports an outcome. A missing outcome, an unknown outcome and a late outcome are all execution coverage, not part of the decision:
- No replay or re-execution. OpenProdkt does not run your call, so it never replays or re-executes one, whether or not an outcome was reported. A missing or
unknownoutcome is not an instruction to retry. Reconcile with the upstream system and decide yourself. - No change to authorization, approval or billing. An outcome report, a late report and a stored coverage event (
execution.outcome.missing,execution.outcome.not_reportable) never change the decision, an approval's state, or what is billed. - Stored as evidence only. When the operator has turned on outcome deadline monitoring, a missing outcome is stored as a coverage event. It is a record that no report arrived, not an action taken on the call.
See Policy Engine for the ordered evaluation stages, operator and type rules, and worked examples, and Developer Guide for how to handle each outcome in your application.