Jev 1.13
Answers a typed placement question over the candidate ids, a yes/no hold question and a risk score.
81,825 · ±25,834 · 360 ms · #4
Brains
Hosted models answer the same typed question as the local baselines: the board, every legal placement simulated one step ahead, and a request for a probability over those placements. Figures on the right are mean score · ±95% · median latency · rank.
Hosted models
DeepSeek V4 Flash
Reads the rendered board and candidate table, answers with one candidate id in strict JSON.
71,705 · ±19,580 · 1.3 s · #5
GPT-5.6 Luna
Reads the rendered board and candidate table, answers with one candidate id in strict JSON.
41,912 · ±10,133 · 1.2 s · #6
Qwen 3.7 Flash
Reads the rendered board and candidate table, answers with one candidate id in strict JSON.
19,355 · ±7,174 · 760 ms · #7
GPT-4o mini
Reads the rendered board and candidate table, answers with one candidate id in strict JSON.
2,655 · ±488 · 1.4 s · #8
Gemini 3.5 Flash Lite
Reads the rendered board and candidate table, answers with one candidate id in JSON.
not yet run
Claude Haiku 4.5
Reads the rendered board and candidate table, answers with one candidate id in strict JSON.
not yet run
Local baselines
Greedy heuristic
One placement ahead, weighing holes, height, bumpiness and completed lines.
274,207 · ±14,326 · 0.12 ms · #1
Dellacherie-style
A board evaluator in the Dellacherie tradition: landing height, transitions, wells and holes.
269,744 · ±10,020 · 0.12 ms · #2
Two-ply search
Searches the current piece and the next piece with a board evaluator.
267,675 · ±2,019 · 4.0 ms · #3
Random legal
A seeded uniform choice over the legal placements. The floor of the field.
427 · ±25 · 0.17 ms · #9
Every brain is a thin adapter over the same harness. To enter one, see Add a brain.