---
title: Scaffold Harness
emoji: "🧪"
colorFrom: gray
colorTo: green
sdk: gradio
app_file: app.py
pinned: false
license: other
license_name: FSL-1.1-ALv2
license_link: https://github.com/sxc3030-eng/scaffold-harness/blob/main/LICENSE.md
short_description: Paired GAIN/LOSS tests for layers built on LLMs.
---

# scaffold-harness

Measure whether an agent, RAG pipeline, router, verifier, or other layer built
on top of an LLM improves answers or breaks them.

This Space runs only the built-in deterministic smoke comparison. It requires
no model or API key and performs no GPU computation. The free ZeroGPU wrapper
is present only because the selected Hugging Face runtime requires it. The complete public source,
documentation, tests, examples, and FSL-1.1-ALv2 license are available in the
[GitHub repository](https://github.com/sxc3030-eng/scaffold-harness).

The Python adapter imports the callable named in a configuration. Never run a
configuration you did not write or audit yourself.