ARCHITECTURE

Six layers. One source of truth.

AURORA is not a model — it's a substrate. Each module composes the same six layers so what works for glioma generalises to spina bifida, and what works in one hospital generalises to the next.

system architectureMIT @ public alphafederation-first
I · Six layers

The substrate, laid out.

Layer I · Data Layer
aurora.layer.datalayer

Open registries, imaging, omics

Federated registries of imaging, omics, surgical video, follow-up and patient-reported outcomes. Permissive defaults, strict provenance.

ReproducibleFederated-by-defaultProvenance-signedAudit-logged
II · Motivations

Why the substrate looks like this.

Five commitments the architecture is built around. Each one is the answer to a specific failure mode we've seen in hospital AI.

Federation
Data is heavier than code. Move the code.
AURORA-FED · RFC-0001
Provenance
Provenance is a load-bearing element, not an audit checkbox.
Sigstore receipts
Equity
Equity belongs in the loss, not in a slide deck.
RFC-0042 · equity gate
Override
The clinician keeps the override. The system keeps the receipt.
Audit log · v0.4.8
Patient voice
If the patient isn't in the room, the room hasn't started yet.
Two permanent council seats
Five commitments · one substrate
III · Deployment

From a single laptop to a 100-hospital network.

Three reference deployments. The code, the weights and the signatures are identical across all three.

Mode A

Workstation

Single GPU box or Apple-silicon laptop. Useful for local research and one-off analyses. No federation. No PACS integration.

Mode B

Hospital · on-prem

Containerised inside the hospital network. DICOM and FHIR connectors. Federation node optional. Default for pilot sites.

Mode C

Federation

Compute travels to data across a permissioned mesh of hospital nodes. Differential-privacy budgets enforced per registry.

IV · The system in operation

From DICOM to signed report.

AURORA's six layers are a mental model. This is what a single case actually does as it moves through the substrate, and the environments the substrate is tested against. Every hop is observable, every output is signed, and nothing about the path is hidden from the hospital running it.

A request, from DICOM to signed report

Eight hops, all observable, all signed. The numbers in brackets are the SLOs the substrate enforces on a reference workstation; modules inherit and tighten them.

  1. 01
    DICOM arrives [≤20ms]
    C-STORE handler (Rust · dcmtk-rs) writes the study to a temp namespace; nothing is admitted yet.
  2. 02
    Consent + provenance check [≤8ms]
    Substrate verifies cohort policy, Sigstore-signed weight digests and federation key. Mismatch → refuse + audit.
  3. 03
    Case bind [≤12ms]
    Postgres write of the case envelope under the module's tenant; pgvector embedding row reserved.
  4. 04
    Foundation pass [20–60ms]
    PyTorch + Triton Inference Server on the GPU; weights are mmapped, not re-loaded. Module-specific.
  5. 05
    Subsystem fan-out [parallel]
    Asynq task queue dispatches subsystem jobs in parallel; failures retry under a per-job idempotency key.
  6. 06
    Synthesise + render [≤80ms]
    DiagnosticReport synthesised; PDF rendered with Typst; DICOM SR emitted; URLs are hospital-local.
  7. 07
    Audit chain close [≤15ms]
    NDJSON audit envelope signed with the site's federation key; hash chain extends; durable before the next render.
  8. 08
    Deliver [≤30ms]
    Outputs flow back to PACS (STOW-RS) and EHR (FHIR R4). Override marker primed for the MDT surface.

Compatibility matrix

What we test against on every release. "Supported" means a passing CI run on dedicated hardware; "best-effort" means a community run we accept but do not gate on.

Lane
Supported
Best-effort
Not supported
GPUs
NVIDIA Hopper · Ada · Ampere
AMD MI300 (via ROCm)
Older than Turing
Linux distros
Ubuntu 22.04 / 24.04 · RHEL 9
Debian 12 · SUSE 15
CentOS Stream 8 and earlier
Kubernetes
1.28 → 1.30
1.27 · vanilla k3s
OpenShift < 4.14
Browsers
Chrome / Edge ≥ 122 · Firefox ≥ 124
Safari 17.4+
IE · pre-Chromium Edge
PACS vendors
GE · Sectra · Agfa · Visage
Carestream · Philips IntelliSpace
Proprietary append-only stores w/o DICOMweb
EHRs
Epic (FHIR R4) · Cerner Oracle · MEDITECH Expanse
OpenMRS · Bahmni
Closed-protocol legacy EHRs
IdPs
Microsoft AD FS · Okta · Keycloak
Google Workspace · Auth0
Shared service accounts
Air-gap
OCI tarball + offline Helm
Manual sneakernet w/ Sigstore verification
Out-of-band manual model patches
V · Regulation & HIPAA

HIPAA is a property of the substrate.

AURORA is decision-support and research infrastructure, not a marketed medical device. Where it touches Protected Health Information it does so as a Business Associate, on the hospital's own estate. Here is exactly how that works — at the substrate, and module by module.

Privacy Rule

Minimum necessary, by construction

AURORA never asks for more PHI than a subsystem needs. Identifiers in DICOM headers and free text are stripped at ingest; models receive de-identified tensors, not charts. The covered entity remains the data steward — AURORA is a Business Associate, not a new owner of the record.

Security Rule

Administrative, physical & technical safeguards

All three required safeguard families are implemented at the substrate: documented access policy and workforce roles (administrative), deployment inside the hospital's own physically-controlled estate (physical), and encryption, unique user identity and audit controls (technical).

Breach Notification

Detect, contain, report

The hash-chained audit log makes unauthorised access detectable and tamper-evident. AURORA surfaces the forensic trail; the covered entity's existing breach-notification workflow (within 60 days, per §164.404) is the binding legal mechanism. AURORA does not relocate that duty.

Business Associate Agreement

A BAA before a single byte

A signed BAA governs every pilot deployment, scoping permitted uses, mandating safeguards, and requiring return-or-destruction of any PHI on termination. Because AURORA runs on-prem and federates compute, in most deployments no PHI is ever disclosed to AURORA at all.

The three Security-Rule safeguards

§164.308–312 require administrative, physical and technical safeguards. AURORA implements all three at the substrate, so every module inherits them.

Administrative§164.308
  • Documented risk analysis per module before deployment
  • Role-based workforce access enforced in policy (OPA / Rego)
  • Per-clinician identity on every override and read
  • Sanction policy and incident-response runbook shipped with the chart
Physical§164.310
  • Runs inside the hospital's own physically-secured data centre
  • No AURORA-operated cloud holds PHI in any pilot deployment
  • Air-gap bundle for sites that refuse outbound connectivity
  • Device & media controls inherit the site's existing facility plan
Technical§164.312
  • TLS 1.3 in transit; site-encrypted volumes at rest (AURORA refuses to render onto unencrypted storage)
  • Unique workload identity (SPIFFE/SPIRE) and mTLS between every service
  • Hash-chained, signed audit log — tamper-evident integrity controls
  • Automatic logoff, session scoping and encryption keys in the site's Vault/KMS/HSM

What every module inherits

Before any disease-specific handling, every module is bound by the same substrate-level controls.

No PHI leaves the site
Federation moves compute to data. In on-prem and federated modes, PHI is never disclosed to AURORA — the strongest possible HIPAA posture.
De-identification
Safe Harbor by default (all 18 identifiers removed); Expert Determination where a cohort needs richer fields, documented per §164.514.
Encryption
TLS 1.3 in transit; the runtime refuses to render onto unencrypted storage at rest.
Access control
Unique per-clinician identity via the hospital IdP (OIDC); RBAC enforced in OPA; no shared service accounts in the PHI path.
Audit controls
Hash-chained, per-site-signed NDJSON log of every read, override and output — tamper-evident and queryable for breach investigation.
Right to withdraw
Patient withdrawals propagate through cohort policies; the site's existing legal flow (HIPAA, GDPR Art. 17, India DPDP) is the binding mechanism.

HIPAA, module by module

Every module handles a different slice of the chart. This is the PHI each one touches and the specific controls applied to it.

In active build
GliomaAURORA-GLIO live
PHI touched
MRI/CT DICOM, molecular pathology & genomic reports, operative notes, service dates, age.
Minimum necessary
The foundation model receives de-identified pixel and omics tensors only; the 18 HIPAA identifiers in the DICOM header are stripped at C-STORE ingest and never reach the GPU.
Technical safeguard
Free-text pathology and operative notes pass a clinical NER scrubber (Safe Harbor) before any model reads them; molecular reports are tokenised to codes.
Retention & withdrawal
Derivatives are keyed to a study UID, not a patient name. A withdrawal request purges the UID across Postgres and object storage and invalidates downstream caches.
HydrocephalusAURORA-HYDRO live
PHI touched
Serial imaging, shunt-series device data, ICU vitals streams, longitudinal service dates.
Minimum necessary
Lifetime surveillance uses a pseudonymous patient key; the re-identification map stays in the hospital IdP and never enters the substrate.
Technical safeguard
Time-series vitals are date-shifted by a per-patient offset held only at the site, preserving intervals without exposing real dates (Safe Harbor §164.514).
Retention & withdrawal
Longitudinal record keyed to the pseudonym; withdrawal severs the IdP mapping and purges derivatives.
EncephaloceleAURORA-ENCEPH beta
PHI touched
Craniofacial CT, 3D surface scans and clinical photographs — facial images are PHI and potentially biometric.
Minimum necessary
Photographs are processed inside the runtime only; just non-identifying morphometric landmark vectors persist — never an identifiable face on durable storage.
Technical safeguard
Surface meshes are reduced to landmark coordinates at ingest; raw photographs are held in volatile memory and discarded after feature extraction.
Retention & withdrawal
Only de-identified mesh derivatives are retained; withdrawal purges them. Defect-specific consent for LMIC partner programmes handled per the site BAA.
CraniosynostosisAURORA-CRANIO live
PHI touched
Infant clinical photographs, 3D head scans, genomic markers, growth/service dates.
Minimum necessary
Photo-first triage runs on-device: the image is scored locally and discarded with no cloud round-trip; only the score and morphometric features persist.
Technical safeguard
Genomic markers are handled GINA-aware and segregated from the clinical surface; family-history fields carry their own access policy.
Retention & withdrawal
De-identified features keyed to a study UID; withdrawal purges the UID and any cached photographic frames.
Chiari MalformationAURORA-CHIARI beta
PHI touched
Cine-MRI flow series, posterior-fossa morphometry, symptom diaries, service dates, age.
Minimum necessary
Flow analysis consumes phase images only; patient-entered symptom diaries are de-identified before the decompression-vs-watch model reads them.
Technical safeguard
Symptom diaries are scrubbed of names, dates and free-text identifiers on capture; longitudinal surveillance keyed to a pseudonym.
Retention & withdrawal
Study-UID keyed derivatives and diaries; withdrawal purges across Postgres and object storage.
Planned — controls specified, not yet deployed
Spina BifidaAURORA-SPINA planned
PHI touched
Fetal and maternal imaging, gestational and service dates, prenatal records — among the most sensitive PHI in any chart.
Minimum necessary
Maternal identity and the fetal record are linked only inside the site; the substrate sees a study-scoped envelope with no maternal demographics.
Technical safeguard
Prenatal records carry a heightened cohort policy; dual-subject consent (parent and future child) is modelled explicitly as first-class metadata.
Retention & withdrawal
Pseudonymous study key; withdrawal honours the site's existing fetal-record retention rules. Module is planned — these controls are specified, not yet deployed.
Tethered CordAURORA-TC planned
PHI touched
Spine MRI, urodynamic studies, neuro-exam notes, service dates, age.
Minimum necessary
The signal and phenotype models consume de-identified imaging and structured urodynamic measures; identifiers are stripped at ingest.
Technical safeguard
Paediatric records flagged for the site's stricter minor-consent handling; free-text notes de-identified before clustering.
Retention & withdrawal
Study-UID keyed derivatives; withdrawal purges across stores. Module is planned — controls are specified, not yet deployed.
Dandy-WalkerAURORA-DWM planned
PHI touched
Prenatal and paediatric posterior-fossa imaging, gestational/service dates, genetics reports.
Minimum necessary
Origin and flow models consume de-identified imaging volumes; genetics reports are tokenised to variant codes.
Technical safeguard
Rare-disease re-identification risk is managed with k-anonymity checks before any cohort export; small-cell suppression on reports.
Retention & withdrawal
Study-UID keyed; withdrawal purges across stores. Module is planned — controls are specified, not yet deployed.
Arachnoid CystsAURORA-ARACH planned
PHI touched
Brain MRI/CT, cyst morphometry, neuro-exam notes, service dates, age.
Minimum necessary
Physics and triage models consume de-identified imaging only; identifiers stripped at ingest.
Technical safeguard
Incidental-finding records carry a consent policy distinct from symptomatic ones; free-text de-identified before modelling.
Retention & withdrawal
Study-UID keyed derivatives; withdrawal purges across stores. Module is planned — controls are specified, not yet deployed.

AURORA is not a substitute for a covered entity's own compliance programme, and nothing here is legal advice. Specific modules destined for clinical use are validated and submitted to the relevant regulator by the deploying institution. The BAA, not this page, is the binding document.

Changelogv0.6.0