CareOS 🏥

Return time to care — without making clinical information less trustworthy.

CareOS is a clinician-first context, interoperability and assurance layer that sits beside existing hospital systems. It composes source-linked patient context and makes that context safely usable by people and bounded AI applications.

Clinical review demo → · One patient end to end → · Synthetic study →

Research / synthetic evaluation only · not for clinical use · no production write-back

The problem

Healthcare workers still act as human middleware between KIS/EHR, LIS, RIS/PACS, documents, ePA, calls and messages.

CareOS explores a stable layer above those systems rather than replacing them:

KIS / EHR · LIS · RIS/PACS · documents · ePA
                       ↓
                reusable adapters
                       ↓
          source-linked clinical context
 identity · provenance · time · lifecycle · freshness
                       ↓
       role-specific UX + bounded agents/apps
                       ↓
                    human authority

The north star is Time Returned to Care — safety gated.

What to try

1. Clinical review — inspect changed, pending and critical context with sources attached.
2. One patient journey — follow source → draft → outage → correction → recovery → human review.
3. Synthetic paired-study surface — see how the project intends to measure time, errors and verification behaviour.

The intended question is not “is this a nice AI demo?” It is:

Can clinicians spend less time hunting, reconciling and copying information without losing provenance, uncertainty or control?

Four correctness invariants

1. Pending ≠ negative.
2. Unavailable ≠ absent.
3. Documented therapy ≠ AI recommendation.
4. Agent draft ≠ source truth.

The model may propose structure. It does not become the authority that creates trusted clinical truth.

What is implemented

• source-linked clinical lifecycle and provenance
• patient-local graph + stale-artifact invalidation
• bounded agent/tool authority and adversarial scenarios
• NORMAL / DEGRADED / OFFLINE / RECOVERY states
• patient/family source-linked experience
• interactive human clinical review UX
• FHIR R4 research read path
• ISiK/FHIR-oriented validation path
• HL7 v2 ADT/ORU parsing for synthetic/deidentified inputs
• hospital manifest / preflight / review-pack tooling
• Docker/Helm deployment scaffold
• synthetic Time Returned to Care study machinery

Agent safety model

untrusted model proposal
        ↓
deterministic Agent Gateway
patient · encounter · task · tools · operations · budgets
        ↓
trusted Tool Proxy
        ↓
source-linked result
        ↓
untrusted draft
        ↓
human review

The model cannot grant itself a new patient, tool, operation or effect authority.

The blocker we do not hide

The frozen 500-case synthetic clinical-truth holdout currently preserves precision/provenance but reaches only 26.32% recall with 100% review-case burden.

Therefore production G1 remains blocked.

That means CareOS should not be presented as clinically validated, production-ready or ready to make patient-care decisions. The next step is better external evidence, not stronger copy.

Real-world proof ladder

synthetic tests
   ↓
real clinicians on synthetic cases
   ↓
independent clinical / privacy / security / IT critique
   ↓
real non-secret hospital capability map
   ↓
approved vendor / deidentified sandbox
   ↓
governed shadow workflow
   ↓
bounded read-only pilot
   ↓
second hospital / second vendor

The first workflow to prove is physician morning review + documentation preparation.

A speed gain is not a win if source verification, error rate or safety gets worse.

Integration truth

Path	Current state
FHIR R4 read	implemented research runtime
ISiK / FHIR read	runtime + validation path
HL7 v2 ADT/ORU parsing	synthetic/deidentified connector
Real HL7 v2 transport	external evidence required
Vendor API	contract only
Live transactional/write	blocked by release policy

No named KIS/LIS compatibility claim is made without a real approved environment.

What is not proven

CareOS is not currently:

• clinically validated
• approved for identifiable patient-data production use
• proven to save clinician time
• proven interoperable with a named production KIS/LIS
• proven repeatable across hospitals
• regulatory approved/certified
• a 24/7 clinical service

Deep review

• Release assurance
• Current status and gaps
• Claim → evidence matrix
• Production gates
• Agent security model
• Real-world proof campaign
• Clinical study protocol

---

Keep systems of record. Standardize trustworthy context above them.

Models may interpret and propose. Evidence, authority and safety boundaries remain outside the model.

Built by Michael Ninh in Berlin.