# Measure the boundary, not a broad security score.

Match the threat to a measurable test.

> **Research guidance.** This guide describes how we think about the problem. It is not a description of a product capability. Where it mentions Vraelis Contour, it says what exists in the experiment and what is proposed.

## Test against a defined threat

A useful evaluation separates model tampering, data poisoning, manipulated inputs and unauthorized actions. Define attacker access, normal operating conditions and the evidence needed to reproduce each result.

Prompt injection and adversarial perception inputs need different test cases and controls. Constrain authority separately from detection.

## Measure the engineering tradeoffs

Measure missed attacks, false alarms, latency, compute use and investigation effort. An attack test result applies to its tested conditions, not to every model or physical environment.

## Distinguish change from compromise

Performance drift can have operational causes. Use qualified baselines before attributing an anomaly to an attack.

These are research areas. Vraelis does not claim universal attack detection and publishes no detection or accuracy numbers.

## How Vraelis Contour is evaluated today

The evaluation is narrow. At the runtime rejection-diagnostics revision, 133 of 133 internal tests pass in a pinned, network-disabled harness, and internal automated testing reproduced that on a separate build on October 10, 2026 (Pacific). That measures automated behaviour inside the experiment. A latency study and live fault runs are not built.

## Read the underlying guidance

[NIST adversarial machine learning taxonomy](https://www.nist.gov/publications/adversarial-machine-learning-taxonomy-and-terminology-attacks-and-mitigations-0)

---

Vraelis documentation, Research. https://docs.vraelis.com/ai-security-evaluation