SentraMind

Primer

Deterministic first, then anything else

Why a governed AI-security decision starts with a byte cap, a parse, and a deterministic detector before any model-shaped step.

A byte cap comes before the parse

The hot path begins by measuring the input. Content past the byte cap is not parsed, so later steps never receive an unbounded body. The cap is the first control a reader can retell without a diagram, and it is the reason a huge paste cannot skip the rest of the design by sheer size.

The claim stops at admission. The cap decides whether parsing starts. Verdict p95 latency is null in this POC, so this primer does not attach a speed to the cap, and it does not promote a handful of development requests into a published figure. Null stays null.

Policy runs on the body the cap admitted

After the body is inside the cap and has been parsed, the path applies the compiled policy snapshot. That snapshot is the decision basis for the request. The action returned to the caller is the action the snapshot produces, and the effective-policy viewer is meant to read the same snapshot through the same evaluator.

A private second policy would make the retelling false. Dry-run, the viewer, and enforcement are one engine on purpose. When a demonstration shows the viewer allowing a body that enforcement blocks, the demonstration has split the engine, and the repair is to put both callers back on one snapshot.

The deterministic tier can complete the action

The tier after policy is deterministic, and in this build it is a single RE2 scan. Patterns that require backreferences or lookbehind are rejected by the lint the runtime enforces, so those constructs do not survive as a style preference the scanner ignores. The same bytes under the same snapshot produce the same action.

Repeatability is the property being named. This POC publishes no precision and no recall for the tier, and it holds no labelled corpus from a live estate against which those figures could be computed. A stable match on a fixture can still be the wrong match for a document the fixture never resembled. Stability and fitness are different sentences, and only the first is claimed here.

A generative model is kept off this path

The action is finished by the cap, the snapshot, and the deterministic tier. A general or generative model is not called to settle cases the tier finds awkward. If it were, the viewer could no longer replay the action from the snapshot alone, and the decision would depend on a system this path is built to exclude.

An encoder, when one is configured, does not take the policy’s place. A score from that tier is not the verdict. When the encoder is absent, the absence remains a visible state, and the path does not invent a substitute model to fill the gap. This POC also publishes no quality figure for the encoder tier, so a score must not be narrated as a measured result.

Later labels stay on their own work

This primer describes the hot path available to evaluate with Platform Core and GenAI Prompt DLP. AI usage observability is coming next, and the other crew directions sit on the design-partner roadmap until evidence and an owner decision promote them. Discussion of those labels can sit beside a verdict. It does not enter the action the verdict returned.

Stand-ins that share the screen stay labelled as well. Officer keys are software, hardware tokens are absent, machine certificates come from a software development CA, and endpoint inventory is a simulated fleet. A deterministic result on synthetic text leaves every one of those labels where it was.

How to read this

How to read this
Read it asWhich meansDo not read it as
Byte cap, then policy, then the deterministic tier Same input, same snapshot, same action A model score standing in for the verdict
RE2 single scan Backreferences and lookbehind are rejected A style guide the runtime ignores
Missing encoder The absence stays visible A silent substitute model

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