This video shows my Tetris test of Lev. Lev is an experimental, Jev-inspired classifier that combines lexical matching signals in a confidence-gated loop. Here, I use its default heuristic backend to choose among all legal placements. This backend runs without neural-network weights, and its probabilities in this test are uncalibrated.

Download the video (.mp4).

Lev’s default heuristic backend reached 47.39% accuracy (11,847 of 25,000 reviews) on the official IMDb test set, below the 50% baseline for this balanced dataset. I used the same negative and positive labels as in my Laya and CLM evaluations, without training or calibration. Classification and result logging took 8.32 seconds on the DGX Spark’s CPU.

For comparison, see my Jev, CLM, and Laya Tetris tests. I discuss the broader text classification context in Language Models for Text Classification: From Bag-of-Words to Jev.