Put ground truth in the knowledge layer

"Can we trust the answers?" starts in the wrong place. Trust is whether your organization can point to the paragraph that justified a coverage decision, a routing rule, or a patient instruction, and prove it was current when the answer was generated. That capability is the Knowledge layer: your truth layer, separate from the rented brain that reads it.

The fine-tune fantasy

The usual AI strategy in healthcare: export a pile of PDFs, fine-tune something, call it "trained on us." Policies change quarterly. Payer bulletins land on Fridays. Weights are stale the day after you ship them. You cannot cite page 14 of a neural network in a deposition.

Ground truth (utilization rules, benefits, clinical pathways, contracts) is owned by the business, served to the model at runtime, governed like any controlled document set.

Three buckets executives should demand

Delegates confuse lakes, models, and tools. Put three buckets on one slide:

1. Knowledge (static-ish truth): policies and playbooks with effective dates, ACLs, and citations in the UI 2. Tools (transactional truth): live eligibility, appointment slots, claim status 3. Model (interpretive work): language and synthesis, with the system of record elsewhere

The Model provides reasoning. Knowledge provides facts you can refresh without a training project. System messages say how to use both. Tools fetch what no PDF should pretend to know live.

Cite or silence

Regulated-facing workflows need an operating rule your quality team can enforce: cite or silence. Without a source chunk and version metadata, the system escalates. Unglamorous. Effective in depositions.

Investment implications

Shift budget from "Do we need a custom model?" toward corpus owners, clinical informatics time, indexing discipline, refresh cadence, and retrieval evaluation.

The counterargument: "Frontier models know medicine." They know patterns in training data. They do not know your Medicaid carve-out for this county on March 1. Medical fluency is different from your operational truth.

Patient intake on the Approach page retrieves specializations and payer rules. That is the pattern: intelligence systems separate reasoning from authoritative facts.

Delegate deliverable: pick one corpus (prior auth policy is a strong first candidate), name an executive owner, define refresh SLAs before the next model bake-off.

Teams that try to memorize policies inside weights usually learn that refresh belongs in the corpus. Building the layer from governed documents is the Implementation follow-on; Strategy is naming an owner and SLA first.