Analysis of learned representations typically focuses on similarity, measuring how closely embeddings align with external references. However, similarity reveals only what is represented, not whether that structure is robust.

Shesha is a framework for measuring geometric stability, a distinct dimension that quantifies how reliably representational geometry holds under perturbation. Across 2,463 configurations in seven domains, research shows that stability and similarity are empirically uncorrelated ($\rho \approx 0.01$). This distinction makes Shesha a necessary complement to similarity for auditing representations across biological and computational systems.

🚀 Quick Links

📄 Foundation Paper: arXiv:2601.09173
📄 LLM Steering and Drift Paper: arXiv:2604.17698
📄 Bio Paper: arXiv:2604.16642
💻 Code: GitHub Repository
📦 PyPI: shesha-geometry

📦 Installation

pip install shesha-geometry

Key Applications

Geometric stability provides actionable insights across multiple domains:

Safety Monitoring: Acts as a functional geometric canary to detect structural drift nearly 2$\times$ more sensitively than CKA.
Controllability: Supervised stability predicts linear steerability with high correlation ($\rho = 0.89$-$0.96$).
Model Selection: Dissociates from transferability, revealing the "geometric tax" that transfer optimization incurs.
Scientific Analysis: Predicts CRISPR perturbation coherence and neural-behavioral coupling.

Citation

If you use Shesha or geometric stability in your research, please cite:

bibtex

@software{shesha2026,
  title = {Shesha: Self-Consistency Metrics for Representational Stability},
  author = {Raju, Prashant C.},
  year = {2026},
  publisher = {Zenodo},
  doi = {10.5281/zenodo.18227453},
  url = {https://doi.org/10.5281/zenodo.18227453},
  copyright = {MIT License}
}

@article{raju2026geometric,
  title = {Geometric Stability: The Missing Axis of Representations},
  author = {Raju, Prashant C.},
  journal = {arXiv preprint arXiv:2601.09173},
  year = {2026}
}

If you use the supervised variants (supervised_alignment, lda_stability, variance_ratio, class_separation_ratio), please also cite:

bibtex

@inproceedings{raju2026canary,
  title = {The Geometric Canary: Predicting Steerability and Detecting Drift via Representational Stability},
  author = {Raju, Prashant C.},
  booktitle= {ICML Mechanistic Interpretability Workshop},
  year= {2026}
}

If you use the shesha.bio module, please also cite:

bibtex

@article{raju2026crispr,
  title = {Directional coherence and effect magnitude in single-cell {CRISPR} perturbation responses},
  author = {Raju, Prashant C.},
  journal = {arXiv preprint arXiv:2604.16642},
  year = {2026}
}

pcr2120/shesha-geometry

Shesha: Self-Consistency Metrics for Representational Stability

Overview

🚀 Quick Links

📦 Installation

Key Applications

Citation

pcr2120/
shesha-geometry