SCIENCE

The methods, in plain view.

Six methodological commitments shared by every disease module. Foundation models on their own are brittle. Mechanistic simulators on their own are blind. AURORA insists on both — with ethics auditing inside the training loop.

foundation modelsPDEcausal inferencefederation
I · Commitments

What we will, and will not, do.

01

Foundation models

  • Vision-language-omics transformers pre-trained on permissively-licensed neuro corpora
  • Patient-conditioned heads per disease module
  • Distilled to clinic-edge sizes for on-prem deploy
02

Mechanistic simulation

  • PDE solvers in Rust/CUDA for diffusion-reaction and CSF compartment models
  • Differentiable physics where useful for inverse problems
  • Tight coupling to data through learned priors
03

Causal inference

  • Counterfactual graphs over treatments, resection extents and progression
  • Trial-aware exclusion of confounded comparisons
  • Bayesian uncertainty surfaced in every recommendation
04

Federated training

  • Compute travels to data, never the reverse
  • Differential privacy budgets enforced per registry
  • Provenance-signed updates across sites
05

Provenance & signing

  • Sigstore-signed releases for every model and benchmark
  • Pinned weights, pinned containers, pinned data hashes
  • Every output traces back to the commit and dataset that made it
06

Ethics-in-the-loop

  • Equity stratifications evaluated inside the training loop
  • Releases failing equity thresholds are blocked, not warned
  • Consent metadata is a first-class object, not an afterthought
II · Layered stack

The methods, top to bottom.

LAYER 5SurfaceDICOM viewer · Trial console · Patient-facing summaries
LAYER 4ReasoningFoundation models · Causal graphs · Mechanistic simulators
LAYER 3RepresentationsDisease-specific embeddings · Cross-modal fusion · Outcome priors
LAYER 2SubstrateFederation runtime · Provenance · Sigstore signing · Audit logs
LAYER 1DataImaging · Omics · Surgical video · Outcomes · Registries
III · Model families

Three model families, one substrate.

Foundation models, mechanistic simulators and causal graphs are co-designed and share the same provenance pipeline.

Foundation models

Per-module vision-language-omics transformers, pre-trained on permissively-licensed neuroimaging, surgical-video, and biomedical literature corpora. Distilled to clinic-edge sizes so they run on hospital hardware. Conditioned on patient-level signals — never on identity.

Mechanistic simulators

PDE solvers in Rust and CUDA for diffusion-reaction, CSF compartment, and biomechanical problems. Differentiable where useful for inverse problems. Tightly coupled to the foundation models through learned priors so the two stay in conversation, not in opposition.

Causal graphs

Counterfactual graphs over treatments, resection extents and progression. Trial-aware exclusion of confounded comparisons; Bayesian uncertainty surfaced in every recommendation. The clinician sees the band, not just the number.

IV · Limitations

What AURORA is not.

AURORA is research and decision-support infrastructure. It is not a medical device, it is not a regulator-approved diagnostic, and it is not a replacement for clinical judgement. Modules destined for clinical use are validated and submitted to regulators by partner institutions — the underlying code remains open.

Federated learning does not eliminate privacy risk. We treat it as one tool among several: differential-privacy budgets per registry, signed updates, and audit logs are the others. Equity audits inside the training loop reduce, but do not erase, the harms of biased data. We publish failures alongside successes.

Finally: not every recommendation a model produces should be acted on. Every output explains itself. Every override is logged. The clinician is the override.

Changelogv0.6.0