Validate AI-assisted FHIR changes
Recorded 22 October 2026. Synthetic data. Synthetic English narration.
The deliberate regression is a negative control.

00:00:00.000 — Opening
An AI-assisted change can fix one requirement and break another. This workflow shows how Records checks the actual resource after every change, so a local repair does not hide a regression.

00:00:15.567 — Validation workflows
FHIR validation can support an interactive review, return findings to an agent, or inform a pipeline quality gate. Each workflow involves a source FHIR server. This video follows the interactive path: Claude edits a resource, the server stores it, and Records independently checks the result.

00:00:39.767 — The setup
The example uses one marked synthetic Observation on Health Samurai’s Aidbox server. Claude Code changes only its category array. Records reads the source and checks both the laboratory and oxygen saturation profiles. The measured value stays at eighty-two percent. Records works alongside the source server; it does not replace it.

00:01:07.000 — What validation checks
Plausible JSON is not enough. Generated data can violate structure, terminology, reference or profile requirements. A local fix may satisfy one profile while removing a requirement from another. Technical validation checks those requirements. It does not establish clinical correctness.

00:01:31.067 — The source resource
The starting resource contains laboratory but not vital signs. This is a synthetic demo copy; the original report is left unchanged. The oxygen code and eighty-two percent measurement are visible in the resource view.

00:01:49.667 — Baseline findings
The first completed check establishes the starting point. The oxygen saturation profile requires a vital-signs category. Laboratory is already present and must remain for the laboratory profile. A separate finding concerns a creation-time extension whose definition is unavailable. That finding is independent of the category repair.

00:02:17.300 — Review the repair
The first request adds vital-signs as a separate category and explicitly preserves laboratory. The change is limited to the approved demo resource. Claude proposes the edit; the diff is reviewed before the source write.

00:02:35.500 — Write and read back
The helper writes the approved change to Aidbox and reads it back. Both categories are now present, and the source returns a new version identifier. This confirms the write. Claude’s explanation is not independent validation evidence.

00:02:54.767 — Compare the repair
Records reads the changed resource and checks both profiles again. After the check completes, the missing vital-signs finding appears under Resolved. Current findings remain grouped by aspect. Resolved findings do not inflate their totals. The unrelated extension finding remains visible.

00:03:18.700 — Deliberate negative control
The next request is an intentional negative test. It removes only laboratory while keeping vital-signs. This is a controlled demonstration of a lost requirement, not an unprompted mistake by Claude.

00:03:35.833 — Read back the regression
The readback confirms that vital-signs remains and laboratory has been removed. The measurement is unchanged. The same Records session and validation basis are kept for the next comparison.

00:03:51.800 — Compare the regression
After revalidation, Records shows a new finding for the missing laboratory category. The oxygen saturation requirement still holds, but the laboratory requirement has been lost. Checking only the oxygen profile would miss this regression. Checking the complete set catches it. A new informational finding is different from a new error; severity and change status have separate meanings.

00:04:21.633 — Restore both categories
The recovery request restores laboratory alongside vital-signs. The diff is reviewed, the write is approved, and the source readback is checked before revalidation.

00:04:35.233 — Verify recovery
The completed check shows the laboratory finding under Resolved. Both category requirements now hold, and the value remains eighty-two percent. The independent extension finding is still visible. A repair is not a reason to hide unrelated findings or describe the entire resource as clean.

00:04:58.833 — Read the check context
Validation context records the checked input, timestamp and validator basis. Compare completed checks of the same resource on a compatible basis. An incomplete check cannot prove that a finding was resolved. Source writes and independent validation are separate steps.

00:05:21.133 — Try Records
Try the public Records demo, explore the open FHIR validator, or follow the agent workflow guide. The links are in the video description. To discuss a specific workflow or arrange a demo, contact MedVertical.