Add kp-detect-v18 submission

[kprofundis]

kp-detect-v18 — Kareem Elsamadicy (Independent Researcher)

Release date: 2026-05-25

Detector: kp-detect-v18

Summary

A meta-stacker built on top of v15 (PR #155). On the 339K RAID test rows where v15 was uncertain (score in [0.2, 0.8]), this submission refines predictions using a small bag of physics-grounded features (Mahalanobis distance to per-class meaning centroids, semantic phase-space velocity/acceleration, conceptual density, and first-order coherence at lag 10). The refined scores are blended with v15 base at alpha=0.95 (95% v15 + 5% physics-meta). v15 confident rows are unchanged.

Train OOF (5-fold, balanced n=27,900)

Variant	AUC	log_loss
v15 base alone	0.9870	0.5587
v18 raw blend	0.9899	0.4930

Expected lift on RAID test: modest +0.003 AUROC over v15 baseline (96.44 → ~96.7), with paraphrase-cell AUC lifting on 6/8 domains where v15 was weakest.

Contact: kelsamadicy@gmail.com

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

kp-detect-v18

Release date: 2026-05-25

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 96.46 and a TPR of 89.74% at FPR=5% and 80.89% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 97.53 and a TPR of 92.81% at FPR=5% and 84.97% 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!