README.md

---
title: AI Model X-Ray
emoji: 🔬
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.33.0
app_file: app.py
pinned: false
license: apache-2.0
short_description: Which layers can you prune? Structural health scanner.
tags:
  - backyard-ai
  - custom-ui
---

# 🔬 AI Model X-Ray — Structural Health Scanner

Which layers of your transformer are compressible? Which are fragile?

Select a model or paste any HuggingFace model ID. The scanner extracts
attention graphs, computes the spectral simplicial hierarchy per layer,
and classifies each layer as **immune** (safe to prune), **buffer** (caution),
or **critical** (do not touch).

Based on the spectral principle λ₂(T(G)) ≤ λ₂(G), validated on 45,000+
graphs with zero violations. Formally verified in Lean 4.

## How it works

For each layer we average every attention head's map over a small probe set
(16 sentences for text models, 16 CIFAR-10 images for vision), flatten it to a
signature, and join heads whose signatures correlate (Pearson r > 0.3). On that
head-to-head graph `G` we compute:

- **λ₂(G)** — algebraic connectivity of the head graph.
- **T(G)** — the triangle graph (edges of `G` that share a triangle).
- **λ₂(T(G))** and the coherence ratio **ρ = λ₂(T(G)) / λ₂(G)**.
- **FI** — the fragility index: fraction of edges sitting in zero triangles.

High ρ with zero FI means a layer is triangle-redundant — its head structure has
slack and is safe to prune. Low ρ means the layer is structurally load-bearing.

| Regime | Condition | Meaning |
|---|---|---|
| 🟢 Immune | ρ > 0.8, FI = 0 | safe to prune |
| 🟡 Buffer | 0.5 ≤ ρ ≤ 0.8 | prune with caution |
| 🔴 Critical | ρ < 0.5 | do not prune |

Pre-loaded models (BERT, GPT-2, ViT) show **instantly** from a precomputed
cache. Custom model IDs and DistilBERT trigger a live scan on ZeroGPU.

🎬 Demo: [YouTube link TBD]
🐙 See also: [Octopus AI](https://huggingface.co/spaces/build-small-hackathon/octopus-ai)

Built by [Cognitive Engineering](https://cognitive-engineering.dev) 🇨🇭