The record can be current and still be wrong

A connected record feels authoritative because it arrived from a health system or wearable moments ago. The transfer can be current while the fact is wrong. A medication may remain listed after it was stopped. A diagnosis may be provisional. A device may misclassify activity. Two systems may disagree. A field may be missing because the source never captured it. Synchronization establishes recency of transfer, not truth of content.

OpenAI's Health guidance tells users that connected information may be incomplete or out of date, gives the example of a medication remaining after discontinuation, and recommends checking important details against the original source. This is a valuable disclosure. The operating system around the data should make those limitations actionable inside the answer.

Every consequential fact needs provenance: source organization or device, record type, observation time, transfer time, status, and any user correction. The answer should distinguish a clinician-entered prescription, a pharmacy fill, a self-reported medication, and a model inference. Flattening them into takes medication removes the evidence needed to resolve conflict.

Preserve disagreement and correction

When sources conflict, the system should not silently choose the newest or most specific field. It can show the disagreement, ask for clarification, and recommend verification with the appropriate professional for important decisions. A coherent summary is not always the safest summary; sometimes the disagreement is the most important fact.

User corrections need careful status. A person may know that a medication stopped yesterday before the formal record updates. Their correction should influence the conversation and remain visibly attributed, not overwrite the upstream source as though it had always agreed. Later synchronization can confirm, complicate, or reverse the correction.

Freshness tells you when the evidence arrived; provenance tells you what kind of claim arrived.

Derived trends need their own evidence. Comparing lab results across institutions can involve different reference ranges, units, methods, and clinical contexts. Wearable metrics can change algorithms. A trend line that ignores these changes can create precision without comparability. Preserve the measurement conditions that make comparison valid.

Deletion and disconnection should not create a false historical record. If source data is removed while prior conversations remain, the system should be able to explain that the answer used information no longer connected. It should not continue presenting the old inference as a current fact merely because the conversation persists.

Freshness rules should follow the fact, not the connector. Demographics may remain stable for years, medications can change in hours, wearable measurements may arrive continuously, and appointment status can change after one phone call. A single last synced timestamp applied to the entire record gives false comfort. Define review windows and warnings by information type and consequence of error.

Missingness should remain visible. Absence of a diagnosis, medication, allergy, or result in one connected source is not evidence that none exists. Coverage can vary by provider, date range, data standard, and patient-matching quality. Summaries should say which sources and periods were available and avoid universal claims from a partial record. A completeness indicator can be more useful than another confident sentence.

Correction workflows should return value to the person. When the assistant identifies a likely discrepancy, it can help formulate a question for the source organization, point to the original entry, and preserve a temporary user annotation while formal correction proceeds. It should not claim to amend the clinical record unless the source confirms that action. Evidence quality improves when the boundary between conversational correction and source correction stays explicit.

For important decisions, present verification as part of the task rather than a generic disclaimer at the end. Identify the specific fact that may be stale or conflicting, show its source and date, and explain what original system or professional can confirm it. Targeted verification helps the user act; repeated boilerplate can satisfy a warning requirement while leaving the actual uncertainty buried in an authoritative summary.

Operate the boundary

The practical starting point is a named control for source provenance, observation and transfer times, conflict state, user correction, and comparability for connected health facts. Write the boundary in terms an operator can evaluate: the initiating principal, permitted purpose, affected resources, allowed consequences, escalation path, expiry condition, and evidence produced. A policy sentence is useful context; the enforced object and its observable state are what make the policy operational.

Test the boundary by introducing stale medications, conflicting sources, changed units, device-algorithm updates, missing fields, and user corrections across synchronized conversations. Preserve the starting state, the agent's route, any intervention, the final effect, and the gaps in observation. Repeat the exercise after changing a model, tool, provider, policy, or data source. A control that passed once should not silently lend its assurance to a materially different system.

The leading signal is important health claims with visible provenance and freshness, unresolved conflicts surfaced, and corrections that remain attributed rather than silently overwriting history. Pair it with a consequence measure so teams do not optimize the dashboard while weakening the outcome. Review both on a fixed cadence and after every material incident or migration. When the signal disappears, determine whether the risk disappeared or the instrumentation did.

The health information and clinical-safety owners should own the decision to continue, narrow, pause, or expand the workflow. The owner needs authority over the control and access to its evidence; responsibility without either becomes ceremonial. Record the decision, the evidence cutoff, the residual uncertainty, and the next review date so the claim can age honestly.

— Dispatches · Summit Cognitive