Pilot model runs
DiffusionGemma: 48/60 tasks, 57/70 questions. Jev: 51/60 tasks, 60/70 questions.
CLM direct replay: 23/60 tasks, 30/70 questions. All 62 choice questions have null option descriptions, which CLM embeds as bare option keys. Two post-hoc input-format checks scored 26/60 and 21/60; the direct score remains primary.
GPT-6 Luna (medium reasoning, strict JSON output): 60/60 tasks, 70/70 questions. Its generative API and latency are not directly comparable to the System One services.
Decider-4B v2 (pinned v2 checkpoint, native typed API): 48/60 tasks, 56/70 questions.
Decider-2B-Vision: 35/60 tasks, 44/70 questions.
GLiNER2.5-Decide: 21/60 tasks, 29/70 questions.
Paired outcomes: 45 both, 3 DiffusionGemma only, 6 Jev only, 6 neither.
These are one-shot results on a selected, openly keyed set. They do not rank models generally. See EVALUATION_PROTOCOL.md for collection and scoring details.