Five seconds to orient
Changed · pending · contradictory · stale · unavailable · current source-linked context · work that needs a human.
Living ground truth · 18 Aug 2026
What should healthcare feel like when information finally works?
Not “more AI.” A healthcare operating layer where people stop acting as human middleware between fragmented systems — and clinicians, patients, hospitals and bounded agents share trustworthy context without losing provenance, privacy or human judgment.
Current boundary: CareOS is synthetic / pre-hospital research. The future journeys and time targets below are hypotheses to test, not clinical outcome claims.
North star
Open the patient once. Know what changed. Know what is uncertain. Know where it came from.
Hospital #100 inherits hospitals #1–99
Reusable adapters, capability profiles, conformance tests and compatibility evidence — not another bespoke integration project.
legacy KIS · LIS · RIS/PACS · ePA · documents
↓
reusable adapters
↓
source-linked canonical clinical context
identity · provenance · time · lifecycle · freshness
↓
deterministic policy + audit
↙ ↘
human interfaces bounded agents/apps
↘ ↙
human authority
Every stakeholder
One foundation. Different superpowers.
The interface should adapt to the job. Nobody gets the entire record by default merely because the platform can retrieve it.
Physician
- Changed since last review
- Pending/conflicting context
- Source-linked timeline
- Prepared note/discharge draft
- Original evidence one click away
Nursing
- Shift changes
- Do next / overdue
- Pending tasks/results
- Care-relevant exceptions
- Calm handover, not more alerts
Emergency
- Critical minimum context in seconds
- Allergy/medication/recent events
- Recency + provenance visible
- Governed break-glass
- No dependence on AI for core truth
Pharmacy
- Medication reconciliation
- Changes/discontinuations
- Allergy + relevant context
- Conflicts prepared for review
- No silent resolution
Diagnostics
- Indication + relevant prior context
- Preliminary/final/corrected lifecycle
- Critical-result acknowledgement
- Images stay in imaging infrastructure
- Context links, not needless copies
Discharge / case mgmt
- Need visible earlier
- Minimum necessary context prepared
- Missing fields explicit
- Structured routing/status
- Less phone/fax chasing
Practice / GP
- Structured referral
- Hospital changes understandable
- Follow-up explicit
- Results return digitally
- No institution-change retyping
Patient / family
- What happened
- What changed
- What is pending + owner
- What happens next
- Ask / flag / share / delegate
Hospital IT
- Manifest + discovery
- Adapter plan + conformance
- Known egress/data flows
- Upgrade preflight
- Canary + rollback
Adoption
It should feel like it was always there.
Nothing clinical depends on CareOS.
Workflow archaeology, baseline measurements, capability manifest, deidentified/synthetic conformance, identity/security review, one bounded workflow.
Shadow mode. Observe reality.
The team works normally. CareOS shows what it would surface or draft. Every mismatch becomes a test, not an excuse.
Fix recurring friction
Clinician champions + implementation + engineering review real failures daily.
Bounded read-only copilot
Only if identity, source state, verification and safety evidence are good enough.
Earn dependency
Expand only when people voluntarily use it and the safety metrics hold.
The test
After a week, the desired reaction is not “the AI is impressive.” It is: “Why was I ever searching five systems to answer this?”
Outcome contract
Return meaningful time to care.
These are product targets, not measured CareOS outcomes. A safety-stop event invalidates a speed win.
| Workflow | First target | Mature ambition | Proof |
|---|---|---|---|
| Physician review + documentation | 20–30 min / shift | 45–60 min / shift | paired time + source checks + errors |
| Nursing handover | 10–15 min / shift | 20–30 min / shift | time + omissions + alert burden |
| Discharge / transfer | 15–20 min / case | 30–60 min / case | time + calls/messages + missing loops |
| Hospital IT routine supported change | days → hours | <1 workday | integration hours + custom code + conformance |
| Patient understanding | baseline | high teach-back | pending + medication + follow-up comprehension |
Synthetic inspiration
Before → after.
Morning review — before
KIS → old note → lab → microbiology → meds → letter → phone → mental timeline → rewrite
Morning review — target
open patient → changed/pending/conflict → inspect consequential sources → decide → edit grounded draft
Discharge — before
search → copy → fax/call → wait → call again → rewrite status → patient still unsure
Discharge — target
minimum context → missing fields → digital route → structured status → next owner → patient sees next step
Critical infrastructure
Design the failure before the feature.
Current sources
Source read/verify + bounded assist. Write still requires a separate capability.
Stale / dependency loss
Show state visibly. Disable absence claims and unsafe actions.
No source/network
Legacy fallback. No fake empty state. No hidden clinical writes.
Network returns
Reconcile missed versions/events before normal actions resume.
Permanent invariants
wrong patient → BLOCK
pending ≠ negative
unavailable ≠ absent
stale ≠ current
agent draft ≠ source truth
read ≠ write
model proposal ≠ authority
translation ≠ source mutation
FHIR-valid ≠ trusted issuer
faster + less trustworthy = FAIL
Clinical context graph
Healthcare is relational and temporal.
The graph is a derived, patient-local map over evidence — not a model-created source of truth.
patient
├─ encounter
│ ├─ diagnosis
│ ├─ order → specimen → result
│ ├─ medication
│ └─ document
├─ pending task
└─ follow-up
Questions it unlocks
- What result belongs to this episode?
- What corrected fact superseded the old one?
- Which draft relied on a fact that changed?
- What evidence supports this sentence?
- Which downstream workflow must reopen?
Agents
Make every user more capable. Never give the model hidden authority.
Orientation agent
Changed / pending / source / context.
Documentation agent
Grounded notes, letters, handovers and forms.
Coordination agent
Missing fields, referrals, replies, owners and follow-up.
Verification agent
Grounding, lifecycle retention, missing evidence.
Patient agent
Explain, translate, find, prepare questions.
IT agent
Explain preflight/conformance failures without granting itself infrastructure authority.
agent identity/version
+ human/workflow delegator
+ patient/encounter/task
+ allowed tools/operations/data
+ egress destinations
+ budgets/timeouts
+ evidence requirements
+ exact human confirmation
+ audit
+ revocation/kill switch
Patients
Access is necessary. Understanding is the product.
Patient home
- Today / next
- What happened
- Medication changes
- Pending result + owner
- Appointments/follow-up
- Documents + original source
- Who accessed/shared data where supported
- Ask / report possible error / share
Boundary
- Purpose and legal access still govern visibility
- Restricted/sensitive data is not exposed just because it is retrievable
- Plain language is presentation, not source mutation
- Children/proxies/guardianship need dedicated policy
- Low-digital-literacy and no-smartphone paths are mandatory
24/7 relationship
If hospitals depend on it, we operate it like infrastructure.
Continuous
Connector health, security advisories, non-PHI fleet telemetry where approved.
Rollout weeks
Short recurring clinician + implementation + engineering review.
Stable operation
Monthly/quarterly value, safety, support and roadmap review.
Every upgrade
Compatibility → conformance → canary/shadow → promote or rollback.
Critical incidents
On-call, severity model, hospital incident channel, post-incident regression.
Flywheel
A generalisable incident at one hospital becomes protection for the next.
Master sequence
Think globally. Prove locally.
Foundation
Truth · identity · graph · audit · resilience · agent authority · adapter/install · measurement.
One irresistible workflow
Infectiology orientation/handover/discharge preparation.
One real hospital
Observe → sandbox → shadow → read-only pilot.
Repeatability
Second ward → second vendor → second hospital without core fork.
Cross-provider
Hospital ↔ practice ↔ post-acute ↔ patient continuity.
Germany → EU → global
National rails, EHDS/IPS, trust, country packs, interoperable ecosystem.
Where are we?
Architecture is already unusually developed; executable pre-hospital foundations are meaningful. But real hospital evidence, repeatability and national operating scale are still near the beginning. That is why the next phase is reality, not feature count.
~45–55% is a rough estimate only for the foundations we can build alone — not “52% of healthcare solved.”
Ground truth
The hard questions stay visible.
Wrong patient. Automation bias. children/proxies. sensitive data. total outage. mass bad update. liability. vendor lock-in above FHIR. who owns conflicts. who can see what. how to leave the platform. whether time saved becomes patient time. We do not bury these in a risk appendix.
Conversation starter
We have pushed the synthetic problem far enough to need reality.
The useful next conversation is not “do you like CareOS?” It is: which assumptions survive a real hospital, which are already solved better, and which unsolved seams are worth building together?