Geometric Assurance Suite

Spectral diagnostics and model integrity tooling for transformer systems

API-accessible diagnostics, research-grade instrumentation, and advanced drift monitoring for teams that need measurable signals about model behavior.

Geometric Assurance is a spectral diagnostics and model integrity suite for transformer systems. Geometric-brain-mcp is the front door, Unitarity-lab is the research engine, and VAR is the advanced monitoring layer.

What the suite does

Geometric Assurance is a spectral diagnostics and model integrity suite for transformer systems.

Product architecture

One front door, one research engine, one advanced monitoring layer.

Front Door

Geometric-brain-mcp

API, MCP, and Python interface for spectral diagnostics, comparisons, and audit workflows.

Open repo

Research Engine

Unitarity-lab

Research and audit engine for deep instrumentation, hidden-state analysis, and reproducible experiments.

Open repo

Enterprise Beta

VAR

Enterprise beta monitoring layer for drift, anomaly, and integrity workflows.

Open repo

Evidence tiers and analysis modes

Hidden-state and eigenvalue analysis are primary evidence-bearing modes. Text proxy mode remains available only as an indirect demo path.

Primary Evidence

Hidden-state analysis

Direct instrumentation of latent activations for model comparison, drift monitoring, and reproducible audits.

Primary Evidence

Eigenvalue analysis

Spectral structure measurements to support anomaly detection and integrity monitoring workflows.

Indirect / demo-only / low-evidence mode

Text proxy mode

Text proxy mode is a lightweight demo and screening path. It does not inspect true hidden states and should not be treated as equivalent to hidden-state or eigenvalue analysis.

Audit workflow

A practical engagement path designed for trust and fast first outcomes.

1. Submit repo, model, or output samples

Share the target components and examples for initial scoping.

2. Run diagnostics

Execute spectral diagnostics and model comparison passes.

3. Receive findings and recommendations

Get a structured summary of measured signals, caveats, and next actions.

4. Expand into deeper instrumentation or monitoring

Extend into research-grade instrumentation or pilot monitoring based on goals.

Sample output / report

Sample audit report coming next.

  • Scope summary and diagnostic inputs.
  • Model comparison tables and drift snapshots.
  • Hidden-state and eigenvalue analysis notes.
  • Anomaly and integrity observations with confidence qualifiers.
  • Recommended follow-up instrumentation and monitoring actions.

Self-Serve Audit Hub

Educational research-station workflow for spectral structure measurement and model comparison.

Measures spectral structure, not output correctness; does not detect hallucination or judge content quality.

How to connect your HuggingFace model

Three paths from any HuggingFace model to a spectral audit. The notebook is the fastest start.

  1. Load your model in the starter notebook — Click Launch Gemma 4 Starter Notebook below. The notebook imports any HuggingFace model via AutoModelForCausalLM.from_pretrained("your/model-id"), attaches activation hooks, and computes hidden-state eigenvalues automatically.
  2. Run the spectral measurement — The notebook computes the GUE r-statistic from eigenvalue spacings of the activation Laplacian. Use the external proxy below (opt-in) to submit text for a remote spectral estimate.
  3. Interpret your numbers — Use the Foundational Methods card below to interpret your r-statistic. For production-grade hidden-state and eigenvalue analysis, follow the links in the Deeper Analysis section.
Launch Gemma 4 Starter Notebook
Research-grade educational workflow. Local spectral analysis runs in-browser. External text proxy is disabled by default and requires explicit opt-in.
Launch Gemma 4 Starter Notebook

External proxy returns r-ratio and regime from spectral analysis. For production work, use hidden-state and eigenvalue analysis in the research engine.

Foundational Methods

Key reference parameters for spectral rigidity measurement.

Spectral Rigidity

GUE r-statistic

r ≈ 0.5996

The GUE (Gaussian Unitary Ensemble) r-statistic measures the average ratio of consecutive eigenvalue gaps in the model's hidden-state weight matrices. The Wigner surmise predicts r ≈ 0.5996 (empirically ~0.602); Poisson-distributed spacings give r ≈ 0.3863. This measures spectral structure only — it does not indicate output quality or coherence.

  • r near 0.5996: Eigenvalue spacing matches GUE prediction — normal spectral rigidity for this regime.
  • r near 0.3863 (Poisson): Eigenvalue repulsion breaks down — possible layer collapse, weight degeneracy, or uncorrelated blocks. Investigate with hidden-state analysis.
  • r > 0.65: Stronger-than-expected eigenvalue repulsion — possible over-regularization, rank deficiency, or unusual training dynamics.

Deeper Analysis

Text proxy mode is a screening shortcut only. For production-grade results, connect directly to the research engine or the API layer.

Research Engine

Unitarity-lab

Full hidden-state and eigenvalue instrumentation. Attach activation hooks, run reproducible audit notebooks, export audit trails. This is where r-statistics and spectral metrics are derived from real hidden states rather than text proxies.

Open research engine →

API + MCP Front Door

Geometric Brain MCP

REST API, MCP server, and Python SDK for spectral diagnostics and model comparison workflows. Integrate into CI pipelines, compare model checkpoints, or run batch audits across a model family without custom instrumentation.

Get API access →

Ecosystem Paths

Two core strategies for audit and monitoring integration.

Front Door

Geometric Brain

API and Python interface for lightweight spectral diagnostics, model comparisons, and audit initialization. Integrate with your pipeline for ongoing screening.

Learn more

Enterprise Beta

VAR Monitoring

Advanced drift, anomaly, and integrity monitoring layer. Pilot-ready for teams wanting operational drift detection and tamper-evident controls.

Learn more