---
title: "KYE™ at the inference boundary (NVIDIA NIM · vLLM) — authority before the model runs · KYE Protocol™"
description: "A KYE™ Policy Enforcement Point that an OpenAI-compatible inference gateway — NVIDIA NIM, Triton, vLLM, TGI — consults before running the model. Every agent inference call is authority-checked at the last point before action: admit, block, escalate, or quarantine. It makes Open Access Authority™ enforced, not just declared."
url: https://kyeprotocol.com/nvidia-nim/
lang: en
source: "KYE Protocol"
---

> A KYE™ Policy Enforcement Point that an OpenAI-compatible inference gateway — NVIDIA NIM, Triton, vLLM, TGI — consults before running the model. Every agent inference call is authority-checked at the last point before action: admit, block, escalate, or quarantine. It makes Open Access Authority™ enforced, not just declared.

# Authority before the model runs

The inference call is the last point before an agent acts. **KYE™** puts a Policy Enforcement Point there: an OpenAI-compatible inference gateway — NIM, Triton, vLLM, TGI — consults it before running the model, and the answer is _admit_, _block_, _escalate_, or _quarantine_.

This is what makes [Open Access Authority™](https://kyeprotocol.com/open-access-authority/) _enforced_ rather than merely declared — authority is checked at the physical boundary where the action would happen.

[Start a governed pilot](https://kyeprotocol.com/poc/) [See Open Access Authority™](https://kyeprotocol.com/open-access-authority/)

## How it works

This is infrastructure you deploy, not an app you open: a thin PEP sits in front of your inference endpoint.

1. **Extract.** From each request, derive the proposed action — invoking a model, or, if the request declares tools, invoking a tool (the consequential case).
2. **Decide.** Consult the central KYE™ admissibility decision for this agent, right now — reusing the same decision engine the rest of the protocol uses.
3. **Enforce.** Map the verdict to a gateway action: proceed, block with a 403, escalate to approval, or quarantine the agent.
4. **Evidence.** Seal the decision into an [Evidence Pack™](https://kyeprotocol.com/evidence-pack/), replay-verifiable from public keys alone.

## Why the inference boundary

For a platform or security team, this is the one place enforcement cannot be bypassed.

- It is the last gate before action: every agent path, whatever framework, funnels through inference.
- A request declaring tools — a wire transfer, a database write — tightens the check exactly when it matters.
- The EU AI Act's logging and human-oversight duties bind at runtime; this is where you discharge them.

Boundary: KYE™ governs whether the agent may act; it does not run the model or proxy the tokens — that stays with your inference engine. Part of the [KYE Protocol™](https://kyeprotocol.com/) platform; pairs with the [Weights & Biases provenance integration](https://kyeprotocol.com/weights-and-biases/).