Splash is an LLM analysis framework built on Structured Wave Geometry (SWG) and Structured Coherence Geometry (SCG). It provides tools to analyze, visualize, and regulate coherence in large language models by projecting hidden states into a geometric space of invariants.
Modern LLMs are powerful but fragile. They drift, hallucinate, loop, or collapse into incoherent states. Splash tackles this by bringing geometry-based diagnostics to LLMs:
- Invariant measures — entropy, curvature, alignment, dislocation, flux.
- Dashboards & plots — coherence maps, layer curves, dislocation heatmaps.
- Control policies — temperature/top-k regulation, adaptive decoding.
- Calibration & profiling — thresholds tuned for stability and robustness.
- Robustness testing — perturbation and paraphrase analysis.
In short: Splash is how you make coherence visible and actionable in LLMs.
adapters.py— Interfaces for Hugging Face, API clients, and custom models.orchestrate.py— High-level entry points for evaluating prompts and hidden states.types.py— Core dataclasses for hidden states, invariants, and evaluation runs.scg/— Geometry core (curvature, entropy, flux, projection, dynamics).geometry.py,layers.py,curves.py— Hidden-state analysis functions.dashboard/— Visualization: coherence dashboards, invariant grids, heatmaps.control/— Policy, controller, and actuator modules for runtime LLM regulation.robustness.py— Perturbation testing (synonyms, paraphrases, distractors).roi_miner.py— Region-of-interest mining for episodes and traces.train_dynamic.py— Training routines for dynamic adapters.
Splash is designed as a research and monitoring framework for LLM coherence:
- Detect collapse fingerprints in LLM hidden states.
- Visualize how coherence evolves across layers and tokens.
- Run dashboards that highlight when an LLM is drifting or looping.
- Apply control policies that intervene in decoding when collapse is imminent.
- Compare robustness across prompts, perturbations, or models.
Other domains (physics, biology) provided inspiration, but Splash’s focus is language models.
git clone https://github.com/whelmed/splash.git
cd splash
pip install -e .