| --- |
| title: LLM-X-RAY Leaderboard |
| emoji: 🔬 |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 5.20.0 |
| app_file: app.py |
| pinned: true |
| license: mit |
| short_description: Geodesic LLM Auditing & Hallucination Observatory |
| tags: |
| - leaderboard |
| - spectral-analysis |
| - hso-perceptron |
| --- |
| |
| # LLM-X-RAY: Operator-Geometric Auditing & Hallucination Observatory |
|
|
| This Space implements the official **HSO-Perceptron (Hilbert-Schmidt Operator)** benchmark and auditing registry for Large Language Models. |
|
|
| ### Architecture |
| - **Layer A (Weight SVD):** Stable rank ($r_s$), effective rank ($r_{\text{eff}}$), and condition number ($\kappa$). |
| - **Layer B (Activation / HSO Observer):** Ledoit-Wolf regularized covariance, $\tau = 0.95$ observable subspace rank ($d$), and blind quotient manifold ($\ker$). |
| - **Layer C (Behavioral Battery):** Factual probes, adversarial traps, and paraphrase stability. |
