Grounding AI Content in Sources It Can Point To
A real, code-enforced pipeline for source-grounded AI content already exists in our backend — extraction, verification, and citation, most of it handled by code rather than a model's word alone.
"AI content is generic and made up" is a fair criticism of a model asked to write from nothing. The answer to it isn't a bigger disclaimer — it's a pipeline that makes the model show its work, and checks that work with code rather than trusting it. That pipeline is real and built, inside the ContentOS backend, even though it isn't something a reader can point at from the content app today.
The eight steps
The design, as documented directly in the backend's own source: Request → Intent → Retrieval → Source Ranking → Evidence Pack → LLM → Claim Extraction → Claim Verification → Output.
What's notable about that sequence is how little of it is actually the model's job. Six of the eight steps are ordinary code: understanding what's being asked, retrieving relevant material, ranking sources by authority, assembling an evidence pack from what was retrieved, and — critically — verifying whatever claims come out the other end. The model only sits at one step directly (generating from the evidence pack) and proposes at another (extracting claims from its own output) — and both of those model-touched steps get checked by code immediately afterward, not taken at face value.
Why the verification step is the part that matters
Any system can retrieve some documents and stuff them into a prompt — that's the easy half of "grounded" content. The harder, more valuable half is checking afterward that what the model actually wrote is supported by what it was given, rather than assuming retrieval alone guarantees accuracy. Claim extraction and claim verification exist specifically to catch the gap between "the model had good sources" and "the model's output actually reflects them."
Where this sits today
There's no public route into this pipeline yet, and nothing in the content app's own interface calls it. We're describing it now because the code is real, tested, and already reflects a considered answer to a real objection — not because there's a feature waiting behind a hidden setting. When source-grounded generation becomes something you can use directly, this eight-step design is the mechanism it'll be built on.