Can Jev pass
AI 101?

Behaves like a sensible reflex agent: cleans dirt and waits when both rooms are clean. It sees both rooms while the textbook rule sees one, and a three-line rule with the same view behaves the same way.

88% clean with 3.7 moves, vs 87% and 15.7 moves for the rule (mean of 10 seeded runs).

Luna, reasoning off · 88% clean, 9.1 moves

Weighs many soft signals into one probability you can probe: single-change tests show what moves each call. It leans slightly towards leaving compared with the 1995 tree.

Agrees with the AIMA tree on 30 of 40 random scenarios (3 of 5 presets); waits in 22 vs the tree's 24. Neither is ground truth.

Luna, reasoning off · agrees with the tree on 31 of 40 random scenarios

Code computes the exact risks; Jev picks the move. It dies more often than the hand-written agent: comparing numeric risks is a documented weak spot, and routine moves are often near coin-flips. The hand-written agent also plans across the whole known map; Jev sees its neighbours.

Escaped with gold in 23 of 50 worlds vs 29 for the hand-written agent; 10 deaths vs 3. Mean score 245.5 vs 512.1.

Luna, reasoning off · gold in 16 of 50 worlds, 1 deaths

With the current state alone it locks into a two-state loop. With a record of visited states it solves the puzzle. With lookahead features it is optimal, but so is a no-model rule given the same features, so at that level the features do the planning.

State only 0/10 · + memory 10/10 · + lookahead 10/10 optimal. Random mover with memory: 98% solved, 33% optimal.

Luna, reasoning off · state only: solved 0/10; with memory 10/10

With the board alone it beats a random player 12 times in 20 but also loses 5, and a perfect opponent beats it. One-move tactics from code make it mostly draw; with the perfect-play result per square it never blunders.

vs perfect play · board only: 4 draws, 16 losses in 20 · + tactics: 14 draws, 6 losses in 20 · oracle: 20 draws, 0 losses in 20.

Luna, reasoning off · board only vs perfect play: 7 of 20 not lost

Two designs. Counting (A* with Jev's distance estimates) is a documented weakness; ranking boards with search in code is TypeSafe's recommended pattern. Jev's rankings carry real information (far above a ranker with none) but neither design beat the one-line Manhattan heuristic.

A*, depth 16: 408 nodes with Jev + features vs 226 with Manhattan; 19% of estimates too high. Ranking with search in code: solved 16/30 vs 29/30 for the Manhattan ranker (a perfect ranker 30, a ranker with no information 1).

Luna, reasoning off · ranking with search in code: solved 12 of 30

Its headline use case: many narrow judgments, each a choice among 49 labels, checked against human annotation. With no training on the corpus it matches an HMM trained on it.

92.9% on 2,593 words vs 92.3% for an HMM tagger and 88.4% for most-frequent-tag (Jev − HMM -0.7 to +1.9 points, 95%).