SentraMind

Feature

One policy engine for the viewer and for enforcement

Dry-run, the effective-policy viewer, and the verdict path call the same deterministic evaluator. If they could disagree, the design would be wrong.

Three callers, one evaluator

Dry-run, the effective-policy viewer, and the verdict path are three callers of one deterministic evaluator. The same snapshot, the same subject, and the same input are supposed to produce the same action in all three. A second engine kept “for the screen” would let the operator read a decision the enforcement process will not make.

The viewer is how an operator reads the decision the verdict path will make. It is connected to the rules the control plane holds. Under the local-password identity used in this POC, a subject picker can be empty, because that provider does not list directory subjects. That is a property of the identity stand-in, and the action the viewer shows is still the one the evaluator returned.

Precedence is part of the single engine

Inside the evaluator the order is fixed. An exception outranks a rollout stage. A rollout stage outranks the rule action. One engine with an undocumented order would be harder to explain than two engines, because an operator could not tell which clause won. The viewer’s trace shows which clause won under the active snapshot. Dry-run shows what a candidate would change, by action and by rule, before anyone moves the pointer.

A rollout can place one group on enforce and leave another on an earlier stage. The viewer has to show that split for the subject in front of it, and enforcement has to apply the same split. If a demonstration needs the two to diverge so that a story lands, the demonstration is wrong and the story should be rewritten around the clause that actually won.

Versions and the pointer

Publishing a policy writes a new version, whose rules no product path changes, and moves the active pointer to it. The verdict process enforces the compiled snapshot that pointer selects, swapped as a whole. Rollback moves the pointer back. An evaluation that “fixes” a rule by writing SQL has stepped outside the product.

Publish, rollback, and exception grants all need a maker and a separate checker, and a self-approval is refused. In this phase those are two local accounts, so the check separates accounts rather than verified people. A grant takes effect only when a publish or rollback compiled after it activates. An exception stops applying at its expiry with no publish, on every path that evaluates policy, including the verdict path, the viewer, and dry-run. Each path reads its own service’s clock, so expiry is no control against someone who holds that clock. Revocation takes one person and does not close the exception at once: it stays enforced until a publish or rollback drafted after the revocation is approved and activated, or until its own expiry.

What a drift test looks like

A useful test authors a rule, replays a synthetic fixture in dry-run, and reads a real change in the per-rule account. It then publishes, moves the relevant group to enforce, and shows the decision change through the policy surface’s own trace. The control-plane half of that story is what the viewer and dry-run share with enforcement’s evaluator.

The enforcement process has its own half: a verdict posted before the publish and the same verdict posted after, on the hot path. Those two halves are both part of calling the design proved for a given fixture. A slide that shows only the viewer has shown one caller. The sentence next to it should say so, and it should name the synthetic input.

Where the promise stops

The promise is agreement between callers on a snapshot. Latency stays unpublished, because no percentile has been measured for the verdict path. A rule can be right and still sit on a roadmap capture path, and the limit in that case is the coverage cell. The engine’s agreement does not pull a roadmap channel forward.

When a snapshot needs the encoder, that tier contributes hits and the evaluator decides, including when the tier is absent and the rule’s fail-open or fail-closed gate applies. The encoder has no private action beside the snapshot. Generative models stay off the verdict path. Agreement between the viewer and enforcement is worth having because both are dull in the same way: same inputs, same action, written down.

How to read this

How to read this
Read it asWhich meansDo not read it as
Callers Dry-run, the viewer, enforcement A second engine for the screen
Order inside it Exception, stage, rule An order that lives only on a slide
A viewer-only edit A defect A preview

Product statements on this page follow the POC maturity ladder. A coming-next or roadmap item is not available for the current evaluation.