Shesha

Pronunciation: SHAY-shuh (/ˈʃeɪʃə/)

Self-consistency metrics for representational stability analysis.

Shesha measures the geometric stability of high-dimensional representations by quantifying the self-consistency of their pairwise distance structure (RDMs) under controlled internal perturbations.

Full documentation at shesha-geometry.readthedocs.io

Installation

pip install shesha-geometry

For single-cell biology workflows (shesha.bio), install with the bio extra, which adds anndata and scikit-learn:

pip install shesha-geometry[bio]

Quick Start

import numpy as np
import shesha

X = np.random.randn(500, 768)  # (n_samples, n_features)

stability = shesha.feature_split(X, n_splits=30, seed=320)
print(f"Feature-split stability: {stability:.3f}")

For the full API reference, installation guide, and usage examples, see the documentation.

Known limitations

• feature_split measures coordinate-axis redundancy, not basis-invariant geometry. Orthogonal rotations can change the score while leaving pairwise Euclidean RDMs unchanged. See What Shesha does not establish.
• sample_split and anchor_stability correlate unmatched distance entries. Their current outputs must not be used for scientific inference; they emit FutureWarning and will be replaced in 0.3.0.

Tutorials

Explore shesha with these interactive notebooks (most take < 5 minutes; the vision tutorial is longer because it loads several pretrained models):

• LLM Embeddings - Analyze embedding stability across layers and models using feature_split.
• Steering Vectors - Compute steering vectors from contrastive pairs and measure their consistency.
• Vision Models - Compare geometric stability across ResNets, ViTs, and other architectures.
• Representational Drift - Measure drift from Gaussian noise injection and LoRA fine-tuning.
• Training Dynamics - Track geometric stability during training to detect representation collapse.
• CRISPR (Bio) - Use shesha.bio to analyze coherence in single-cell CRISPR perturbation experiments.

Citation

If you use shesha-geometry, please cite:

@software{shesha2026,
  title = {Shesha: Self-Consistency Metrics for Representational Stability},
  author = {Raju, Prashant C.},
  year = {2026},
  howpublished = {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:

@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:

@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}
}

License

MIT

_{Logo generated by Nano Banana Pro}