Reconstruct the patient.
- Open KIS + yesterday's note
- Find lab and microbiology
- Check medication and old letter
- Look for infection-control context
- Call to clarify one unresolved item
- Mentally build timeline, then write it again
Synthetic future-day simulator
What changes for the person actually doing the work?
Switch roles. The same source-linked clinical foundation becomes a different interface, a different agent job and a different proof target. These are hypotheses to test—not current CareOS outcomes.
Synthetic only. No live patient data. No clinical efficacy claims.
Open the patient. Context is already there.
- Changed: what became new/different
- Pending: what is not final
- Review: contradictions/uncertainty
- Now: source-linked relevant state
- Work: human plan + prepared draft
- One click opens the original source
Prepare. Never silently decide.
Pre-round synthesis · changed-since-yesterday · timeline · source-linked progress-note skeleton · missing-field check.
model proposal → patient/task/tool policy → source-linked context → reviewable draft → physician decision
Faster but less trustworthy.
Missed pending item · wrong patient · unsupported claim · source verification collapses · stale context feels current · draft is mistaken for treatment recommendation.
Handover is a scavenger hunt.
- Free-text nursing notes
- Medication/order changes
- Isolation/infection-control information
- Separate task list
- Verbal clarification
- Create the next shift's mental model
Show me what changed this shift.
- Shift changes
- Do next / overdue
- Pending samples/tasks
- Care-relevant exceptions
- Conflicting documentation highlighted
- Source available behind important change
Draft the handover, not the care plan.
Extract tasks from approved plans · structure voice draft · group changes · surface missing required documentation. Human confirms governed documentation.
Handover omissions or alert burden rise.
A faster handover is a failure if important work disappears, duplicate tasks increase or nurses need to fight a new notification stream.
Coordinate by chasing.
- Find aftercare need
- Search chart for facts
- Clarify missing details
- Copy into referral
- Phone/fax providers
- Track replies + re-document status
Enter once. Route structurally.
- Aftercare need visible earlier
- Minimum necessary context prepared
- Missing fields explicit
- Digital request through approved rail/network
- Structured status and replies
- Changed source fact reopens affected field
Prepare the package.
Missing-field collection · source-linked referral draft · reply/status summary · follow-up tasks · patient/family explanation. No autonomous placement decision.
Minimum necessary becomes “send everything.”
Data minimisation and purpose-limited views are part of correctness. Inappropriate sharing or missing-info loops invalidate the workflow.
Leave with documents, not understanding.
- Verbal explanation under stress
- Dense discharge letter
- Medication changes hard to compare
- Pending result easy to forget
- Next steps spread across pages
- Call later because plan is unclear
Know what is happening.
- What happened
- What changed
- What is still pending + owner
- What happens next
- Medication changes + source
- Ask / flag possible error
Explain and prepare questions.
Jargon explanation · translation · locate result/document · prepare questions. Always distinguish source record, clinician plan and AI explanation.
The patient becomes more certain than the record.
Pending, uncertain and unavailable information must stay visibly uncertain. No AI reassurance should exceed the source evidence.
Every site rediscovers the same integration.
- Meetings to inventory systems
- Manual vendor/version discovery
- Custom mapping/code
- Hand-built network/security docs
- Upgrade breaks discovered late
- Knowledge stays in people/tickets
Integration is a product.
- Hospital Capability Manifest
- Automatic adapter maturity selection
- FHIR capability discovery
- Conformance + identity gates
- Version compatibility registry
- Upgrade preflight + canary + rollback
Explain blockers, never invent capability.
Generate preflight explanation · compare manifests · summarise conformance failure · prepare security/data-flow artifact. Contract-only adapters remain blocked.
Hospital #10 costs as much as Hospital #1.
If marginal integration effort does not trend toward configuration + conformance, we built consultancy rather than infrastructure.
The discussion we want
Which assumption breaks first in a real hospital?
This is intentionally specific enough to criticise. The next step is not adding another feature—it is comparing these flows with real clinicians, real integration teams and real hospital constraints.