Benchmarks you can inspect.

Task-specific results with frozen inputs, disclosed failures, and downloadable evidence.

FROZEN DATASET · V1

Support routing

100 held-out cases

Compare Jev, Clef, Clef Flash, OpenAI’s Decisions API and an OpenAI structured-output baseline on the same fictional support-routing cases and instructions.

Accuracy, equal-label macro-F1, invalid/error rates, and latency
Observed mismatches and provider failures before interpretation
Frozen JSON, CSV, manifest, model revisions, and run conditions

Open benchmark Read methodology Publication details

FROZEN DATASET · V1

Support urgency

600 held-out cases

Compare the same five models on a yes/no question, is this support message urgent, and see where each model’s probability puts the line.

Missed urgent messages and false alarms at the 0.5 threshold
The threshold each model needs, and what another model’s threshold costs
Every recorded answer and probability, with frozen JSON, CSV and manifest

Open benchmark Read methodology