# Prism-Eval

One canonical identity. Human page: https://www.insightits.com/products/prism-eval.html

| Field | Value |
|-------|-------|
| Slug | `prism-eval` |
| Status | published |
| Family | aiVerification |
| Infra category | AI Verification |
| Record type | product |
| Parent | — |
| JSON | https://www.insightits.com/catalog/prism-eval.json |
| Markdown | https://www.insightits.com/catalog/prism-eval.md |
| GitHub | https://github.com/insightitsGit/prism-eval |
| PyPI | https://pypi.org/project/prism-eval/ |
| Install | `pip install "prism-eval==0.3.0". Apache-2.0, Python 3.10+, framework-agnostic. Soft CTA EVAL — mailto:info@insightits.com?subject=EVAL.` |
| Version | 0.3.0 |

## What it is

Pre-deploy adversarial CI harness. Runs G4 corpora (digit drops, prompt injections, layout/OCR drift) in pytest or CI and fails the build on false accepts — before poisoned tool calls reach money engines.

## Problem

Runtime gates cannot catch what never failed in CI. Extraction agents need a red-team harness that fails the build on false accepts.

## Alternatives

- **Ad-hoc pytest without an adversarial corpus** — No published competitor comparison page. Prism-Eval is the CI half of the Eval + Shield pair.

## Features

- **G4 Adversarial Corpora** — Digit drops, ignore_previous / system_override injections, line-item and layout shifts, OCR and fax noise, plus a legitimate-zero case so a real $0 never false-fails as a digit drop.
- **Attack-Aware Oracle** — Canonical money comparison plus a security oracle that detects obeyed injections and truncations against ground truth — not brittle string equality, and not an LLM judge.
- **Fail-Closed CI Gate** — Exit code tracks suite_passed and the G4 invariant requires zero critical false accepts. Exports JUnit, SARIF, JSON, and a sealed blake2b audit receipt.

## Architecture

Framework-agnostic Python ≥3.10 package. Pair: when CI fails → runtime with Prism-Shield. Soft CTA: EVAL.

## Use cases

- Fail CI on digit-drop false accepts
- Prompt-injection and OCR-drift corpora before deploy
- LangGraph / CrewAI extraction agents that write to money engines

## Benchmarks

No published vendor benchmark for this product. Do not invent metrics.

### Disclosures

- No published vendor benchmark for this product. Do not invent metrics.



## GitHub

https://github.com/insightitsGit/prism-eval

## PyPI

https://pypi.org/project/prism-eval/

## Documentation

- [Interactive demo](https://insightitsgit.github.io/prism-eval/)
- [Prism-Shield (runtime companion)](https://www.insightits.com/products/prism-shield.html)
- [Compose with VectorPrism / Manifest / Shield](https://www.insightits.com/products/prism-eval.html#compose)
- [https://github.com/insightitsGit/prism-eval](https://github.com/insightitsGit/prism-eval)
- [https://pypi.org/project/prism-eval/](https://pypi.org/project/prism-eval/)
- [shieldPypi](https://pypi.org/project/prism-shield/)

## Is not

- A runtime gate (use Prism-Shield)
- PrismGuard
- PrismManifest (library gate)

## Related products

- [prism-shield](./prism-shield.md)
- [prismmanifest](./prismmanifest.md)
- [vectorprism-demo](./vectorprism-demo.md)