About

About this project, how to cite it, disclaimers.

This project was created alongside the Mechanistic Validity framework, as a way of systematizing the discussion of mechanisms in neural networks. It came about gradually, as I found the same recurring questions and philosophical issues while reading papers and conducting my own research.

A major inspiration for this project is the MIB benchmark, which aims to unify discussions and evaluations of MI methods. I believe the field needs something similar for discussions about evidence (Mechanistic Validity) and mechanistic views themselves.

My goal is to collaborate with voices within MI, philosophy, causal inference, and any other related fields, so that the framework encompasses all relevant views. We can then establish a common ground and common vocabulary, allowing for more principled MI work and interdisciplinary discussions.

If you are interested in contributing, please reach out to me via email. Any opinions are valuable, and disagreements are crucial for refining the framework.

Thanks, Elliot

My contact info is available on my personal site: elliottower.ai

Citation

If you use or reference this framework, please cite:

@software{tower2026mechanisticviews,
  author    = {Tower, Elliot},
  title     = {Mechanistic Views: An Atlas of Hidden Commitments and a Realism Criterion for Mechanistic Claims},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.21196311},
  url       = {https://doi.org/10.5281/zenodo.21196311}
}

Full paper (preprint): https://zenodo.org/records/21923080

Disclaimers

This website is intended primarily for educational purposes, expanding on the content from the paper and adding detailed background on the philosophical and formal foundations from which the framework draws.

AI assistance was used for assembling sources and drafting initial site content; all material reviewed and verified by the author.