About reward-lens

reward-lens is built by one person and measured by machinery that does not care about that. The library is a mechanistic interpretability toolkit for reward models, the graders that define what RLHF optimizes: eight kinds of reward signal behind one protocol, with attribution, causal intervention, concept probes, geometry, training-dynamics tracking, and optimization-loop recorders over them. Every measurement returns an Evidence object with uncertainty, provenance, and a trust level computed by gates the caller cannot override.

The code is MIT-licensed at github.com/reward-lens/reward-lens, installable as pip install reward-lens. Questions, bug reports, and collaboration all run through GitHub issues.

How to cite

@misc{nadaf2026rewardlens,
    title         = {reward-lens: A Mechanistic Interpretability Library for Reward Models},
    author        = {Nadaf, Mohammed Suhail B},
    year          = {2026},
    eprint        = {2604.26130},
    archivePrefix = {arXiv},
    url           = {https://arxiv.org/abs/2604.26130},
}