Add ODean v14 — clean hybrid ensemble (672K, 312K unique scores)

[kprofundis]

Adds ODean v14 — clean hybrid ensemble.

Composition:

• v11 forward-DNA HistGradientBoosting (full 672K coverage)
• 8 API instruction judges (deepseek, kimi, openai_mini, openai_4o, xai, cohere, haiku-v2, opus) on adversarial test-outlier subset (~1.5% of rows)
• Per-judge weights derived from cross-validated train AUC, no judge silenced
• v11 dominant weight; API judges provide diversification on rows where the local model over-commits

The forward-DNA + judge-RNA architecture is grounded in Watson-Crick pairing physics. v14 preserves the natural distribution (no clamping) — 312K unique scores across 672K rows.

Author: Kareem Elsamadicy
Contact: kelsamadicy@gmail.com

[github-actions]

It looks like this eval run failed. Please check the workflow logs to see what went wrong, then push a new commit to your PR to rerun the eval.

[liamdugan]

Hey @kprofundis it looks like your metadata file has a "description" key which isn't recognized by our evaluation bot. Please make sure your metadata matches the example template_metadata.json file exactly and doesn't contain any extra keys whatsoever

[kprofundis]

Thanks @liamdugan — caught it in your bot's error message and pushed a fix in 1cd4bed (about 2 min before your comment landed). Metadata now matches template_metadata.json exactly with no extras. The eval is re-running now. Sorry for the noise!

[github-actions]

It looks like this eval run failed. Please check the workflow logs to see what went wrong, then push a new commit to your PR to rerun the eval.

[kprofundis]

Diagnosed the second eval failure (max() arg is an empty sequence in find_threshold) — turned out our predictions had 624K of 672K rows clamped to exactly 0.001 or 0.999, leaving no per-domain spread for the FPR=5% threshold to land on.

Replaced with v14 in 86d9ffa: same v11+API ensemble underneath, no clamping, 312K unique scores across 672K rows, score range 0.000311–0.999739. Distribution mirrors what binoculars / fastdetectgpt look like.

Renamed the folder to ODean-v14_clean since v13's defining feature (Platt + clamp) was the bug. Sorry for the churn — eval should run cleanly this time. Thanks for your patience!

[github-actions]

Eval run succeeded! Link to run: link

Here are the results of the submission(s):

ODean v14 (clean hybrid ensemble)

Release date: 2026-04-29

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 93.05 and a TPR of 81.90% at FPR=5% and 66.00% at FPR=1%.
Without adversarial attacks, it achieved AUROC of 95.47 and a TPR of 88.96% at FPR=5% and 76.42% 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!