Are the payment terms stated? Is there a damages cap? Batch the checklist and list what came back false.

A compliance checklist is a batch of yes/no questions. Run them in one request against the document and let the code list what is missing — then extract the fields the checklist referred to, with offsets, so the reviewer lands on the clause.

In development — the benchmark numbers on this page are measured; the Decide-1 model is in training.

1. 1

   One noul per checklist item

   Batched; the whole checklist is one call.

2. 2

   Extract in the same request

   Amounts, dates, parties, clause spans — with offsets.

3. 3

   Flag the falses

   The reviewer sees the misses and the spans, not the document.

CHECKLIST = {
  "payment_terms":  {"type": "noul", "instructions": "Payment terms are stated."},
  "delivery_date":  {"type": "noul", "instructions": "A delivery date is stated."},
  "damages_cap":    {"type": "noul", "instructions": "Liability is capped."},
  # … up to 20
}
r = client.system_one(state=contract_text, questions=CHECKLIST)
missing = [k for k, a in r.answers.items() if a.noul < 0.5]

• Noul — a calibrated yes/no probability
• Extract — spans and records with offsets

Pattern from oflight.co.jp’s 12-use-case guide. On the invoice-processing suite (100 states × 5 questions) Jev scores 0.756 zero-shot and the fine-tuned open encoder 0.734; the “duplicate invoice” noul is where zero-shot open models fail worst (0.11 vs a 0.89 baseline).

Numbers about Jev are its authors' or the cited source's; numbers from the benchmark are ours, on one harness, with the raw JSON published. Full table.