RadGrants/Grounded drafting

Draft the application — every claim backed by a source.

Pick a program, give the facts you can support, and generate a field-by-field draft. A second verification pass checks each claim against its source — unsupported claims are flagged, not shipped.

Target program

Your sources

Facts the draft may assert from. Prefilled with Sureel’s real NSF SBIR facts as the demo.

[S1]

[S2]

[S3]

[S4]

[S5]

Phase I builds the verification layer and measures whether it answers the research question. Objective 1: a machine-checkable declarative policy schema for healthcare-intake dialogue (delivered: schema + reference clinic-intake policy; measured: coverage on a labeled corpus of real intake turns). Objective 2: the runtime pre-utterance policy verifier plus an entailment-based grounding verifier emitting the tamper-evident provenance record (measured: precision and recall on held-out labeled turns). Objective 3: benchmark the wrapped agent against two baselines — an unguarded LLM agent and a prompt-only-guardrail agent — measuring policy-violation rate, hallucination/ungrounded-answer rate, task-completion rate, and added latency and per-conversation cost. Objective 4: mid-conversation routing that classifies in real time whether a request must be handled by BAA-covered infrastructure or escalated to a human, without leaking what the caller disclosed. The go/no-go gate for Phase II is a defensible measured reduction in violation + ungrounded rates over both baselines at acceptable task-completion and latency.

[S6]

[S7]

[S8]

[S9]