Skip to content

tetris-bench@2 · Provisional measurements

TETRIS
BENCH

Models select placements from shared one-step board outcomes. CR compares play with gravity paused. Blitz separately measures what each adapter can do within 100 ms.

Replays →MethodAdd a brain →

What changed

The original results are withdrawn. They mixed network latency with decision quality, used an order-dependent rating and gave adapters unequal board information. Jev chose the first listed option in all 57 accepted IQ decisions. Read the correction and archived recordings →

Both hosted adapters now pass mirrored-board and reversed-option sanity checks. These check basic behaviour, not playing strength. Diagnostic record →

IQ comparisons

Provisional quick suite, 3 seeds. CR is relative to this field. This small sample cannot establish a winner or a significant difference. These are reproducibility seeds, not a held-out test of general intelligence.

BrainMean IQ scoreMean IQ linesCompleted-call p50 / p95$ / IQ game
Dellacherie-styleLocal baseline498917.0200.96 µs / 287.29 µs$0
Greedy heuristicLocal baseline528517.3180.62 µs / 270.25 µs$0
Jevtypesafe/jev-1.1319397.0364.8 ms / 474.7 ms$0.00816
LLM classifieropenai/gpt-4o-mini22858.01.26 s / 2.07 s$0.02362
Random legalLocal baseline4530.0161.58 µs / 368.13 µs$0
Two-ply searchLocal baseline525118.33.9 ms / 4.8 ms$0

CR and rating intervals are withheld until a full 30-seed suite is available. These are raw validation results, listed alphabetically. Rating details →

Generated 2026-09-20 · 50 piece cap · 200 tick cap

Completion and deadlines

Completed-call latency excludes censored timeouts and input preparation. A small latency number alone does not show reliability. Blitz is a systems stress test and does not contribute to CR.

Brain / modeCompleted callsTimed-out callsDeadline-miss ticksInvalid / stale / errorsCompleted-call p50 / p95
Dellacherie-styleIQ150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 0200.96 µs / 287.29 µs
Dellacherie-styleBlitz150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 0215.25 µs / 326.54 µs
Greedy heuristicIQ150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 0180.62 µs / 270.25 µs
Greedy heuristicBlitz150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 0211.87 µs / 356.75 µs
JevIQ150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 0364.8 ms / 474.7 ms
JevBlitz0 / 297 (0.0%)297 (100.0%)600 / 600 (100.0%)0 / 0 / 0Not measured / Not measured
LLM classifierIQ163 / 163 (100.0%)0 (0.0%)0 / 163 (0.0%)0 / 0 / 01.26 s / 2.07 s
LLM classifierBlitz0 / 292 (0.0%)292 (100.0%)600 / 600 (100.0%)0 / 0 / 0Not measured / Not measured
Random legalIQ70 / 70 (100.0%)0 (0.0%)0 / 70 (0.0%)0 / 0 / 0161.58 µs / 368.13 µs
Random legalBlitz70 / 70 (100.0%)0 (0.0%)0 / 70 (0.0%)0 / 0 / 0234.00 µs / 450.67 µs
Two-ply searchIQ150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 03.9 ms / 4.8 ms
Two-ply searchBlitz150 / 150 (100.0%)0 (0.0%)0 / 150 (0.0%)0 / 0 / 03.8 ms / 4.4 ms

Recent comparisons