![\"AI-assisted](\"https:\/\/ratnaglobaltech.com\/blog\/wp-content\/uploads\/2026\/10\/The-AI-drafts-It-never-signs.png\")
<\/figure>\n\n\n\nWhat the AI plane does<\/h2>\n\n\n\n
The drafting model is MedGemma, pinned to a version and served locally. It does not answer from memory. It answers from what retrieval hands it.<\/p>\n\n\n\n
Retrieval runs over pgvector and is scoped three ways before a single passage is returned: to the tenant and facility, to the patient’s consent and the declared purpose of use, and to the note type and specialty. The AI plane reads the encounter, the problem list, medications and allergies, labs and imaging, hospital protocols and prior signed notes through the same role-based access control a clinician would. If a clinician in that role could not open a document, the model cannot read it either.<\/p>\n\n\n\n
Every passage returned carries a source identifier. The model is prompted to produce structured JSON, not free text, and every clinical claim in the output must carry the span it came from. A claim without a span is not a hallucination we detect later; it is a validation failure the gate catches before anyone reads the draft. The prompt hash and the retrieval set are recorded with the output.<\/p>\n\n\n\n
The output is a draft. It is labelled a draft, it carries its provenance, and it cannot enter the record.<\/p>\n\n\n\n
What the gate checks<\/h2>\n\n\n\n
The gate is a set of rule families, each deterministic, each versioned as configuration and reviewed like code. A sample:<\/p>\n\n\n\n
- \n
- Dose against formulary.<\/strong> Metformin 5000 mg once daily in a discharge summary is blocked. Range, route and frequency are checked against the tenant’s formulary.<\/li>\n\n\n\n
- Allergy cross-check.<\/strong> Amoxicillin ordered with a penicillin allergy on the active list is blocked.<\/li>\n\n\n\n
- Lab plausibility.<\/strong> An HbA1c of 64 where the unit is percent is blocked. Units, ranges and dates are checked.<\/li>\n\n\n\n
- Required sections.<\/strong> A progress note missing the assessment for a Texas Medicaid H2014 claim is blocked, because the claim would be denied and the agency would find out forty-five days later.<\/li>\n\n\n\n
- Coding consistency.<\/strong> A level of care recommendation that the CANS score does not support warns. ICD-10, CPT and level of care are checked together.<\/li>\n\n\n\n
- Source coverage.<\/strong> A draft stating “family history of coronary artery disease” with no source span warns.<\/li>\n\n\n\n
- PHI leakage.<\/strong> A draft that references a lab value from another patient’s chart is blocked.<\/li>\n<\/ul>\n\n\n\n
Three outcomes. Pass proceeds to review. Warn requires the clinician to acknowledge with a typed reason, and the reason, the rule and the reviewer go to the audit ledger. Block cannot be signed until the finding is resolved in the record, and nobody overrides it, including the clinician.<\/p>\n\n\n\n
The model has no identity in the system of record. It cannot acknowledge a warning, override a block, or sign. That is not a policy. It is the absence of a credential.<\/p>\n\n\n\n
![\"clinical](\"https:\/\/ratnaglobaltech.com\/blog\/wp-content\/uploads\/2026\/10\/Coding-consistency.png\")
<\/figure>\n\n\n\nThe human note takes the same path<\/h2>\n\n\n\n
This is the part that surprises people. When a clinician writes a note by hand, with no AI involved, it goes through the same gate. The same dose check, the same allergy check, the same required sections for the same payer.<\/p>\n\n\n\n
That is why the gate is worth building even for the clients who never turn the AI on, which is most of them. The behavioral health agency that needs every note to carry the elements a Medicaid auditor looks for gets that from the gate, with or without a model. The eye care centre that wants every operation note to name the eye, the lens and the lot number gets it from the gate. When one of those clients decides to turn on drafting, the safety net is already there and already trusted, because it has been checking their own clinicians for a year.<\/p>\n\n\n\n
Signature and provenance<\/h2>\n\n\n\n
Only a credentialed human signs. On signature, the note is stored in the system of record with its provenance: the model version if a model was involved, the prompt hash, the retrieval set, the gate result, and every edit the reviewer made to the draft. A
note.signed<\/code> event goes to the backbone, where the audit ledger, the operational loops and the analytics plane pick it up.<\/p>\n\n\n\nSix months later, when someone asks why a note says what it says, the answer is in the record. Not “the AI wrote it”. Which model, which version, from which sources, checked by which rules, changed how, signed by whom.<\/p>\n\n\n\n
How we know it still works<\/h2>\n\n\n\n
A model is not a component you test once. It is a component you measure continuously.<\/p>\n\n\n\n
Every specialty and payer configuration has a golden set: curated encounters with known-correct notes. Every change to the model, the prompt or the retrieval configuration runs the golden set end to end: draft, gate, diff against the reference. Hallucination is measured as claims without a matching span, which is a number, not an impression.<\/p>\n\n\n\n
The release gate is simple. No increase in unsupported claims. No new blocks in the golden set. Latency within budget. If any of those fail, the change does not reach a hospital.<\/p>\n\n\n\n
In production, a sample of signed notes is re-scored weekly. Drift shows up in a chart before it shows up as a complaint from a clinician or a denial from a payer.<\/p>\n\n\n\n
What we do not do<\/h2>\n\n\n\n
We do not let the model write free text into the record. We do not let it read anything the requesting clinician could not read. We do not let a rule live only in a prompt; if a safety behaviour matters, it is a deterministic check that can be written down and tested. We do not ship a model change without running the golden set. And we do not describe any of this as “AI-powered”. It is a drafting aid behind a gate that was there first.<\/p>\n\n\n\n
If a rule cannot be written down and tested, it is not a safety control. It is a hope. Ask any vendor showing you AI documentation to point at the gate. Then ask what happens to a note written without the AI. If the answer is different, the gate is not a gate.<\/p>\n","protected":false},"excerpt":{"rendered":"
- Allergy cross-check.<\/strong> Amoxicillin ordered with a penicillin allergy on the active list is blocked.<\/li>\n\n\n\n