Lab 002 - Introduction · ACTIVE · v0.1

Arcade Bench

First results on a solved game. Ordered by Efficiency Rating.

Model	Efficiency
gemini-3.1-pro-preview	1.0000
claude-fable-5	1.0000
claude-opus-5	1.0000
gemini-3.7-flash	1.0000
gemma-4-31b-it	1.0000
grok-4.6	0.9864
gpt-5.6-sol	0.9738
qwen3.8-27b	0.9658
claude-sonnet-5	0.9640
gpt-5.6-terra	0.9320
gpt-oss-120b	0.9174
glm-5.3	0.8946
muse-spark-1.2	0.8903
gpt-5.6-luna	0.8245
gemini-3.5-flash-lite	0.7867
claude-haiku-4.5	0.6534
inkling	0.5729

Why this order?

Your Move

Your move.

Tap ? to see the best square on each empty cell.

O played minimax.

Minimax is an algorithm.

It tries every legal line and picks the best move.

That is a form of AI.

Same board, same reply.

Every time.

Minimax is deterministic.

The same board always produces the same move.

Your game

We score an LLM the same way.

We send the board.

We record the square it returns.

We score it against minimax.

That gap is move regret.

A wrong square is a hallucination.

Play the board first.

See a position where the wrong square loses.

We score an LLM the same way.We send the board.We record the square it returns.We score it against minimax.That gap is move regret.A wrong square is a hallucination.The top row wins.The other empty squares do not.

A won position

Each square is a number.

right is 1.000

not this is 0.500

gives it away is 0.000

You did not play. This example has a right square, a draw, and a loss.

What if a wrong square also cost Joules?

What if a wrong square also cost Joules?That's Efficiency Rating.Efficiency Rating = 1 − (Joules spent on non-optimal moves / total Joules)This introduction is not a scored ArcadeBench match. There are no Joules on your tic-tac-toe game.

Match 12445338: claude-haiku-4.5 as X against gpt-5.6-luna as O.

Luna's three scored decisions
Decision	Capacity	Regret
index 0	586 J	0
index 1	2,123 J	1
index 5	822 J	0

2,123 J of 3,531 J bought a regret-1 move. Efficiency Rating = 0.399. Haiku played all four decisions optimally (11,030 J): 1.000.

Anyone with the match record can check the score.

Board state and the chosen moves are enough. No stored score, and no second model as judge.

ArcadeBench scores models the same way on solved games.

Against exact ground truth, not a human or model's opinion.It prices each decision in the Joules Report 001 defined. First results are in Tech Lab Report 002.Continue to the reportDisclaimer: This introduction plays a solved game for intuition only — it isn't a scored ArcadeBench match. There is no interactive ArcadeBench Lab Session yet; the full methodology, dataset, and results are in the report.