About

About this project, how to cite it, disclaimers.

Mechanistic Validity grew out of a practical problem: reading a paper that reports an ablation score and trying to decide what it establishes. The answer depends on what kind of claim is being made, what evidence would count, and what the measurement itself is known to do — questions that eight scientific fields have answered for their own claims but that mechanistic interpretability has not yet answered for its own.

The framework collects those answers into a single instrument: 36 criteria across five validity types, a pipeline that runs from a declared description mode to a structured verdict, and a set of audits that apply it to published claims. The goal is not to grade papers but to make explicit what each result licenses, so that follow-up work can target the specific gap rather than repeat the same measurement.

If you are interested in contributing, disagreeing, or extending the framework to new domains, please reach out. Disagreements are especially valuable — several criteria exist because someone objected to an earlier version.

Thanks, Elliot

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

Citation

If you use or reference this framework, please cite:

@software{tower2026mechanisticvalidity,
  author    = {Tower, Elliot},
  title     = {Mechanistic Validity: A Validity Theory and Evidence Standard for Mechanistic Claims},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20478480},
  url       = {https://doi.org/10.5281/zenodo.20478480}
}

Full paper (preprint): https://zenodo.org/records/21913952

Disclaimers

This website is intended primarily for educational purposes, expanding on the content from the paper and adding detailed background on the fields 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.