SentraMind

Explainer

The verdict path does not call a generative model

How a decision is assembled from a byte cap, policy, and the deterministic tier, and why a general model stays off that path.

Three inputs produce the action, and a reader can name them

The caller receives an action assembled from a byte cap, a compiled policy snapshot, and the deterministic tier. The cap bounds the body before parse. The snapshot selects the action. The deterministic tier supplies the matches the snapshot’s rules ask for. Someone who never looks at the table on this page can still repeat that order, and the order is the design.

The account is complete at three inputs, apart from the service’s clock, which decides whether an exception in the snapshot has expired. A demonstration that needs a fourth input to explain the action has left the design. The fourth input this path refuses, for the action itself, is a generative model. There is also no quieter policy running beside the snapshot and adjusting the result after the evaluator has spoken. One snapshot, one action.

The viewer and enforcement read one evaluator

The effective-policy viewer and the enforcement path call the same evaluator. Inside it, precedence is exception, then rollout stage, then the rule action. A person looking at the viewer is looking at the order enforcement will apply. That shared call is what makes the viewer evidence of the decision, rather than a separate drawing of it.

Disagreement is a failed demonstration. If the viewer allows a case that enforcement blocks, or enforcement allows a case the viewer blocks, the build under test is wrong. This explainer has no grade called “close enough.” The repair puts both callers back on the same snapshot. Annotating the screen as merely illustrative would hide the split instead of closing it.

Awkward cases stay on the snapshot

A generative model is not consulted when the deterministic tier is unsure. The awkward prompt is still decided by the snapshot and the tier. Invoking a general model only for the hard prompts would park exactly those prompts on a system the viewer cannot replay from the snapshot. That split is the failure this page is written to keep out of the path.

The encoder seam is a disclosed stand-in, and it sits under a different limit from the generative ban. This POC publishes no precision, recall, or quality figure for the seam. An encoder score is not a verdict. When the seam is absent, the absence stays visible as its own state, and another model is not swapped in to make the panel look populated. The action remains the snapshot’s action either way.

A publish moves the pointer, and an in-flight request finishes where it started

Publishing writes a new policy version and moves the active pointer. Enforcement holds the compiled snapshot that pointer selects and swaps that snapshot as a whole. A request that has already started finishes on the snapshot it started with. A pointer move during that request does not splice two versions into one action.

History stays readable because the previous version remains after the pointer moves. Rollback moves the pointer back to an older version and leaves the published rules as they were. This section states that much so the verdict path and the policy history describe one operation. The acknowledgement rule of the audit chain, which advances cadence only on a matching digest, is a different record and is not re-derived here.

The latency field is null, and null is not a benchmark

Verdict p95 latency is null. The field is empty because a measurement at rated load belongs to a later phase and has not been made. A figure computed from a handful of development requests would be hard to distinguish from a benchmark, so the product leaves the field null instead of publishing the lookalike. Throughput, GPU, and model-quality figures are absent for the same reason. They have not been measured.

Null is also not zero, and it is not a claim that the path is fast. An evaluation that needs a latency budget measures one under its own conditions. This page will not invent the budget. It will also not let the lack of a generative model on the path be described as if that absence were itself a timing result. Speed and architecture are separate sentences, and only the architecture is stated here.

How to read this

How to read this
Read it asWhich meansDo not read it as
Inputs Byte cap, compiled snapshot, deterministic tier A second, quieter policy
Generative model Kept off this path A helper “for hard cases”
Latency Left null until it is measured A number from a few development calls

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