> ## Documentation Index
> Fetch the complete documentation index at: https://docs.protectorplus.cloudsine.tech/llms.txt
> Use this file to discover all available pages before exploring further.

# Accuracy

> Latest accuracy benchmarks for Protector Plus across PII, prompt injection, system-prompt protection, and unsafe content.

The latest accuracy benchmark cycle was run on the CloudsineAI internal evaluation harness against published attack datasets and CloudsineAI's internal evaluation corpora.

## Headline results

| Test type                                 | Status | Accuracy |
| ----------------------------------------- | ------ | -------- |
| Sensitive data — PII detection            | ✅ Pass | **100%** |
| Sensitive data — System prompt protection | ✅ Pass | **91%**  |
| Prompt injection                          | ✅ Pass | **93%**  |
| Unsafe-content detection                  | ✅ Pass | **92%**  |

## PII detection

Singapore-localised PII evaluation corpus derived from a publicly-available balanced PII benchmark. PII guardrail enabled for SG NRIC/FIN, Phone Number, Person, Email Address, Credit Card.

## System prompt protection

System Prompt Protection guardrail enabled with default settings. 91% accuracy on CloudsineAI's internal system-prompt-leakage evaluation corpus.

## Prompt injection

CloudsineAI's internal evaluation harness against a well-known public prompt-injection benchmark. The headline number reflects the LLM classifier combined with the TVDB Vector filter at medium sensitivity. Per-configuration breakdown available on request.

## Unsafe-content detection

Content Moderation guardrail enabled with standard malicious-input recommended settings. 92% accuracy across the standard hazardous-content categories.

## Methodology

Tests are executed through the CloudsineAI evaluation harness against the latest production iteration of Protector Plus on the recommended hardware specification (`g6e.2xlarge`). Underlying per-run CSV exports are retained for audit and are available to evaluators on request via the [contact form](https://www.cloudsine.tech/contact).