Filter retrieved passages before the generative call — and check the answer after it.

Similarity search returns what looks like the question, not what answers it. In a community test on an 81-passage corpus, two thirds of retrieved passages were excluded by four typed questions — and an injected passage that ranked #1 by similarity was caught with p = 0.99.

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

1. 1

   Four nouls per passage

   relevant · usable as evidence · contradicts the query premise · contains instructions to the model. Batched, one call per query.

2. 2

   Order by priority

   Drop injections first, then irrelevance, then contradictions you want surfaced rather than hidden.

3. 3

   Verify after generation

   “This answer is supported by the passages” as a noul over (answer, passages). Below 0.5, regenerate or abstain.

checks = {f"p{i}_{k}": {"type": "noul", "instructions": q}
          for i, passage in enumerate(top_k)
          for k, q in CHECKS.items()}      # relevant / evidence / contradicts / injection
r = client.system_one(state={"query": query, "passages": top_k}, questions=checks)
kept = [p for i, p in enumerate(top_k)
        if r[f"p{i}_injection"].noul < 0.2 and r[f"p{i}_relevant"].noul > 0.6]

• Noul — a calibrated yes/no probability
• Score — a position on your ordered rubric

Community validation (oflight.co.jp): 72-passage set, injected passage ranked #1 by cosine (0.584) caught with injection score 0.99; two thirds of passages excluded across 6 queries.

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.