Leaderboards / Decision /

decision-conformance

decisionready

Protocol checks for System One decision answers, re-implemented from finnhll/jev-eval (MIT): structural invariants, batched vs. separate questions, repeat stability, option-order sensitivity and monotonicity across ordered states.

core tier · greedy-nothink

same as quick; the probe set is small by design · ranked by checks passed
SubjectPassedinvariantsbatched vs separaterepeat stabilityoption ordermonotonic
gev-26b-decidegguf-iq4_xs-mradermacher·llamacpp-cuda-b11515 49/4933/336/66/62/22/2
jev-27b-vlgguf-q3_k_m-iammrduncan·llamacpp-cuda-b11370-jev 49/4933/336/66/62/22/2
clefgguf-q3_k_m-bartowski·llamacpp-cuda-b11515 46/4933/333/66/62/22/2
d1-3bfp16-liquidai·decide-server-d1-torch-cuda-tf5.19.0-torch2.14.1-cu126 49/4933/336/66/62/22/2
d1-3bfp32-liquidai·decide-server-d1-torch-cuda-tf5.19.0-torch2.14.1-cu126 49/4933/336/66/62/22/2
d1-omni-600mfp16-liquidai·decide-server-d1-torch-cuda-tf5.19.0-torch2.14.1-cu126 48/4933/336/66/62/21/2
d1-omni-600mfp32-liquidai·decide-server-d1-torch-cuda-tf5.19.0-torch2.14.1-cu126 48/4933/336/66/62/21/2
layaauto-amp-convai·decide-server-laya-torch-cuda-tf5.19.0-torch2.14.1-cu126 49/4933/336/66/62/22/2
clef-flashbnb-nf4-iammrduncan·decide-server-clef-torch-cuda-tf5.19.0-torch2.14.1-cu126 46/4933/333/66/62/22/2
clef-flashmlx-q8-mlx-community·clef-mlx-mlx0.32.3-mlxvlm0.7.4 47/4933/334/66/62/22/2
layafp32-convai·decide-server-laya-torch-cuda-tf5.19.0-torch2.14.1-cu126 49/4933/336/66/62/22/2
d1-3bfp32-liquidai·decide-server-d1-torch-mps-tf5.19.0-torch2.14.1 49/4933/336/66/62/22/2
d1-3bbf16-liquidai·decide-server-d1-torch-mps-tf5.19.0-torch2.14.1 49/4933/336/66/62/22/2
d1-omni-600mfp16-liquidai·decide-server-d1-torch-mps-tf5.19.0-torch2.14.1 48/4933/336/66/62/21/2
d1-omni-600mfp32-liquidai·decide-server-d1-torch-mps-tf5.19.0-torch2.14.1 48/4933/336/66/62/21/2
layafp32-convai·decide-server-laya-torch-mps-tf5.19.0-torch2.14.1 49/4933/336/66/62/22/2

Conformance is pass/fail per check, so it is listed, not ranked. Failures are named on each run's page.

quick tier · greedy-nothink

every probe once; repeats as listed in probes.json · ranked by checks passed
SubjectPassedinvariantsbatched vs separaterepeat stabilityoption ordermonotonic
jev-latestopaque·typesafe provider-opaqueharness had local changes48/4932/336/66/62/22/2

Conformance is pass/fail per check, so it is listed, not ranked. Failures are named on each run's page.