Grant and program staff may use the drafted prose in a funder report. Funders may make funding decisions based on the report, but the model does not compute the figures or decide eligibility. Clients can be indirectly affected if misleading framing changes how program performance is understood.
The model receives a filled narrative plus receipted scalar display allowlist. Client rows, direct identifiers, SQL, paths, and receipt hashes are excluded. The output is prose that must pass grounding before suppression, pass grounding again after redaction, and receive named human approval.
The benefit is editorial assistance without delegating arithmetic. Plausible nonnumeric misinterpretation, disclosure of a small aggregate before suppression, and provider-side retention are the main harms. The tool cannot prove source-data consent, completeness, or fairness.
The feature may remain available only as an opt-in seam. Organizations must authorize the cloud transfer, select and pin a Bedrock model, review provider logging/retention, and perform the final qualitative review. A path that lets the model create a figure or bypass either grounding gate is prohibited.
Owner sign-off: Chelsea Kelly-Reif, 2026-07-12. Reassess on any model, prompt, provider, data boundary, approval, grounding, or suppression change.