Reasoning & memory·Discipline
Claim QA
A conservative evidence-boundary layer for generated explanations and assessment papers. It checks whether generated statements are supported, weakly supported, contradicted, or unsupported by local evidence — and refuses to hide the answer.
The classifications
The primary entailment route is tasksource/ModernBERT-base-nli. If the model is unavailable, the runtime falls back to deterministic claim checks and explicit warnings.
- Supported. Aligned with local evidence under the configured entailment or heuristic checks.
- Weakly supported. Plausible context exists, but the evidence is not strong enough for confident language.
- Unsupported. No sufficient evidence was found.
- Contradicted. Retrieved evidence conflicts with the generated statement.
Unsupported or contradicted statements downgrade confidence and appear in artifacts. They do not automatically rewrite scoring, mutate learning rules, or block the runtime unless a downstream operator chooses to act.
What claim QA prevents
The layer exists to prevent generated materials from quietly drifting into:
- biological efficacy claims
- safety claims
- clinical readiness claims
- patentability claims
- legal conclusions
Computational assessments remain computational assessments until external validation exists. The reviewer surface is the gate; claim QA makes the boundary visible.
Artifacts
data/results/claim_qa_latest.jsonl
data/results/claim_qa_summary_latest.jsonThese artifacts provide reviewable claim discipline for candidate papers, explanation briefs, and scientific handoff.