drag to rotate
Every benchmark score for a large model carries a hidden question: how much of that score comes from the model actually reasoning, versus having simply seen the answer during pretraining? This 2D simulator makes benchmark contamination physical. A rotating ring of glowing dots represents individual eval questions, each pushed out from the centre by a distance equal to its embedding similarity against the training corpus; a dashed circle is the similarity threshold a decontamination filter would use to flag and discard suspect items before re-scoring. A histogram beneath shows the same similarity distribution split by ground-truth contamination. Dial the true contamination rate, the model's genuine skill, and the detector's threshold, and watch the reported accuracy separate from the decontaminated accuracy in real time — along with the recall/false-positive trade-off that governs every real contamination audit.