MultiScale-Hybrid-e5-humdisp detector submission

[Vigora]

This PR adds the MultiScale-Hybrid-e5-humdisp detector submission to the RAID leaderboard.

This is a follow-up to MultiScale-Hybrid-e5-both (PR #187). It keeps the E5-based semantic-embedding features and token-probability statistics, and changes the projection axis: instead of a difference-of-means axis with the human subspace removed, it uses a Fisher direction whose within-class scatter is inflated by the displacement that adversarial attacks induce in human documents. Input text is normalised (Unicode canonicalisation, removal of invisible characters, reversal of look-alike letters) before encoding.

The detector is trained on unattacked documents from the RAID training split only; it never sees an attacked example during training.

The detector was trained on the RAID training split and should be categorized in the "trained on RAID" section. Method details will be described in a paper currently in preparation.

A companion submission, MultiScale-Hybrid-e5-humdisp-adv, is opened separately; it differs only in that attacked documents are added to its training set.

Submission files: metadata.json, predictions.json

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

MultiScale-Hybrid-e5-humdisp

Release date: 2026-09-04

I've committed detailed results of this detector's performance on the test set to this PR.

On the RAID dataset as a whole (aggregated across all generation models, domains, decoding strategies, repetition penalties, and adversarial attacks), it achieved an AUROC of 98.83 and a TPR of 96.25% at FPR=5% and 89.43% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 99.42 and a TPR of 98.06% at FPR=5% and 93.69% at FPR=1%.

If all looks well, a maintainer will come by soon to merge this PR and your entry/entries will appear on the leaderboard. If you need to make any changes, feel free to push new commits to this PR. Thanks for submitting to RAID!