Touchstone

Gate

GOAL
Prove a model can be trusted with the basics before harder tiers count.
WHAT IT TESTS
Instruction-following, tool calls without hallucination, faithful structured output, error recovery.
HOW IT GRADES
Deterministic single- and multi-step prompts graded mechanically (vitals, tools, fidelity).

Standings — gauntlet benches solved

ModelSolved
qwen3.6-27b22/22
devstral-small-2-251211/11
gemini-2.5-flash-litecloud11/11
gemma-4-26b-a4b-it11/11
gpt-oss-120b11/11
gpt-oss-20b11/11
granite-4.1-30b11/11
granite-4.1-8b11/11
claude-sonnet-4.6cloud11/11
qwen3.5-122b-a10b-cluster11/11
qwen3.5-4b11/11
qwen3.6-35b-a3b11/11
qwen3.6-35b-a3b-or11/11
qwen3.6-35b-bf1611/11
qwen3-next-80b11/11
mistral-small-3.1-24b-instruct-250311/11
qwen3-coder-30b10/11
llama-3.3-70b9/11
qwen3.5-0.8b9/11
qwen3.5-2b9/11
deepseek-r1-distill-qwen-32b7/11
phi-4-reasoning-plus4/11

Benches — green = in gauntlet

gate.fidelity.csvgate.fidelity.enumgate.fidelity.longctxgate.fidelity.no_commasgate.tools.error_recoverygate.tools.multi_selectgate.tools.no_hallucinationgate.tools.result_usegate.tools.single_callgate.vitals.json_validgate.vitals.readygate.fidelity.multistepgate.fidelity.transformgate.vitals.answer_only