Glossary

Evidence status

15 pages

A confirmed record linking two or more systems' representations of the same real-world entity — the same company, the same patient — when no shared identifier…
Keeping retrieved data structurally distinct from instructions in the assembled context, so a model cannot mistake content it retrieved for something it was…
A failure mode where an AI system, unable to distinguish trusted instructions from untrusted retrieved data, acts on an instruction hidden inside content it…
The assembled block of retrieved, resolved data — structured text, a table, or JSON — formatted for injection into an LLM prompt.
A single crosswalk entry translating one field in a connected system's native schema into the domain's canonical model.
A registered, credentialed connection to an external system, including which protocol it exposes and which canonical fields it can answer for.
The process of determining that records in different systems refer to the same real-world thing, using deterministic keys, probabilistic matching, and…
A user who invokes an already-published, approved prompt template without going through authoring or approval themselves.
A user who authors, versions, and submits prompt templates through the approval workflow, distinct from a prompt consumer who only uses what's already…
The record of exactly what data, from which system, grounded which prompt, for whom, and when.
Checking permissions against a live, current source at the moment of a request, rather than relying on a permission model baked into an index and refreshed on…
The share of requests served by a registered, approved template rather than resolved ad-hoc.
A response missing data from one or more failed source systems, presented without any indication that it's incomplete.
A legal or regulatory requirement that specific data never be processed outside a defined boundary — on-premises, a specific VPC, or fully air-gapped.
A pre-defined bundle of which entities, systems, and fields a recurring business task (a QBR summary, a headcount justification) needs, so it doesn't have to…