What it is
Builds AI text detectors that explain each prediction as a sum of named, validated features with token-level evidence, while staying close to the original detector's accuracy. Turn a black-box AI text detector into an auditable one that reports named feature contributions and token-level highlights, and provides a pipeline to convert your own detector the same way.
Key results
- Prediction decomposes exactly: logit = bias + sum(importances)
- Localizes evidence per feature as token-level highlights
- Interpretability is tunable with tau from 0.00 to 0.95 without redownloads
Shows AI text detectors can be made human-auditable with explicit feature evidence while retaining near-black-box accuracy.