What Makes a Guideline Copilot Source-Grounded?
Grounding is a workflow discipline, not a single model feature — it shows up in how citations are stored, displayed and checked.
Article
Grounding as a design constraint, not a marketing word
"Source-grounded" gets used loosely across the industry, often to describe systems that simply retrieve some documents before generating text. We use the term more narrowly: a passage is grounded when the interface can point to the specific approved source content that supports it, and a reviewer can open that source in the same view. If a statement can't be traced that way, it shouldn't be presented as grounded, no matter how confident it sounds.
That constraint shapes the architecture more than it shapes the language model. It means maintaining a curated layer of approved source material, tagging content at a granular enough level that individual statements can be linked to individual passages, and building a display layer that keeps that link visible rather than optional.
Why inline citation beats a reference list
A common shortcut is to attach a list of sources at the end of a generated summary, the way a research paper lists its bibliography. That satisfies a checkbox but does little for a reviewer trying to verify a specific claim quickly. Our approach instead attaches a citation to each individual statement or flagged item, so a pharmacist can check the exact supporting passage next to the exact claim it supports, without cross-referencing.
This inline pattern also makes gaps more obvious. If a sentence in a draft summary has no attached citation, that absence is visible immediately, which is itself useful information — it tells the reviewer that statement needs independent verification rather than resting on the approved source layer.
Curated sources, not open retrieval
Grounding is only as trustworthy as the material being grounded against. Rather than pulling from the open web at generation time, our design draws from a maintained, approved source layer — a curated set of reference material selected and organized for the workflow. This keeps the provenance chain short and inspectable: source in, citation out, nothing generated from an untracked origin.
This is a deliberately conservative choice. It trades the breadth of open retrieval for a shorter, more auditable chain of custody between what the system says and where that statement came from. For a professional workflow, we think that trade is the right one.
Where grounding still needs a human check
Even a well-grounded citation can be applied to the wrong context, or a passage can be technically relevant but clinically inapplicable to a specific case's nuances. Grounding reduces the surface area for a reviewer to check — from 'is any of this true' to 'does this citation actually support this statement, in this context' — but it does not eliminate the need for that check.
That's why the citation layer is built to support fast verification rather than to replace it. The design goal is a reviewer who can confirm or dismiss a generated statement in seconds because the supporting passage is right there, not a reviewer who has to trust the system by default because no path to verification exists.
Insights articles explain how this product and its demonstration are designed. They are not clinical guidance and do not represent guideline recommendations.