Know exactly how well your model
catches PII.

NERVE is a deterministic benchmark that scores a model's ability to detect personally identifiable information — in ordinary prose and in the wire traffic your systems actually send. Same dataset, same numbers, every run.

Download the desktop app Use the command line

No account. Reproducible by design. Built by Sonomos, Inc.

What NERVE is

A benchmark harness, not a black box.

3,480 frozen cases74 entity types across 11 groups — names, contact info, government & financial IDs, health data, secrets, and more.
Prose & wire trafficNatural text plus JSON bodies, query strings, and headers — where real leaks happen.
DeterministicSeeded generation and stored raw predictions mean the same model always earns the same score.
Model-agnosticRegex baselines, GLiNER, HuggingFace token classifiers, Presidio, or an LLM-as-detector.

Two ways to run it

Pick the desktop app for zero setup, or the CLI if you live in a terminal.

Recommended

Desktop app — no terminal

Download, open, and NERVE runs in its own window. Everything happens on your machine.

  1. Download the installer for your OS.
  2. Open NERVE.
  3. Pick a model, click Run, read the scorecard.

A built-in Check for updates button tells you when a newer version or dataset is available, so you're never on a stale build.

Get the app
For developers

Command line

The core is pure Python with zero third-party dependencies.

# install (adds the local ML inference extras)
pip install -e ".[ml]"

# benchmark a model and print a scorecard
nerve bench <model>

# or launch the same UI the app uses
nerve serve   # http://127.0.0.1:8420

Full commands and options are in the README.

What you get

One screen, in plain language — the numbers that decide whether a detector is good enough.

NERVE ScoreA single 0–100 headline that combines detection, precision, and speed.
LeakGuard (recall)What share of real PII the model actually caught — the leaks it stopped.
DiscriminationHow well it avoids over-blocking clean text that only looks sensitive.
SpeedThroughput and latency — whether it's fast enough to run inline, in real time.