Every claim ships with the thing that produced it.

AURORA is not built on an external citations library. Each module publishes its own working papers, the code and weights behind them, and the evaluation script that generated every number. If a result stops replicating, the paper says so — that is the whole point of doing this in the open.

Working papers
5drafts
Open RFCs
4of 8
Cohorts described
1federated
Replications
21 failed

Working papers

One paper per subsystem, written by the team that built it. Filter by module, or read the whole set.

Worked examples of the index format. Titles and statuses show how entries will be presented; none is a published paper yet.

  • AURORA-CORE Federation protocol Circulating

    Aggregate-only egress as a design constraint, not a policy

    Why the node protocol has no endpoint that accepts patient records, what that costs in model quality, and how we measured the cost across four sites.

    Protocol spec Reference node Eval script May 2026
  • AURORA-CORE Provenance Circulating

    Signed provenance chains for federated clinical models

    Every dataset, model and result carries a signature that lets a third party verify which code and weights produced a number, without access to the data.

    Sigstore integration Verification CLI May 2026
  • AURORA-CORE Equity gate In review

    Blocking release on subgroup performance: implementation and consequences

    The gate that stops a model shipping when subgroup metrics fall below threshold, and the three models it has blocked so far.

    Gate spec Blocked-model log June 2026
  • AURORA-GLIO GLIO-MMB Draft

    Molecular marker prediction from routine preoperative MRI

    Retrospective performance for IDH and MGMT status estimation, with the failure cases characterised rather than excluded.

    Weights Model card Eval script June 2026
  • AURORA-GLIO GLIO-TME Draft

    Tumour microenvironment features that survive site transfer

    Which imaging-derived features hold up when the scanner and the population change, and which do not.

    Feature set Transfer results June 2026

Papers are listed here once a draft exists. Subsystems still in design do not get a placeholder entry — an empty row would imply work that hasn't happened.

RFC index

Every governance and design decision is proposed in public before it lands. Anyone can open one; the council reads all of them, and patient advocates vote alongside clinical and technical members.

Example RFCs illustrating the format and the kinds of decision that go through it. No RFC below has been ratified by the council — the states shown are illustrative.

RFC
Proposal
Area
State
RFC-001
Aggregate-only egress for all federated computation
No mode may transmit row-level data, including de-identified rows.
Protocol
Accepted
RFC-004
Two permanent patient-advocate seats with voting rights
Observer status was rejected as insufficient during review.
Governance
Accepted
RFC-009
Equity gate blocks release on subgroup failure
Applies even when aggregate metrics meet the target.
Ethics
Accepted
RFC-012
Model cards mandatory before any module leaves alpha
Includes intended use, populations, failure modes, and do-not-use conditions.
Quality
Accepted
RFC-017
Publish replication failures within 30 days of confirmation
Removes the discretion that lets inconvenient results sit unpublished.
Research
In review
RFC-021
Patient-facing guides require clinician sign-off before public alpha
Currently drafts are labelled pending review; this makes sign-off a gate.
Content
In review
RFC-023
Withdrawal procedure for degraded models
Defines who can trigger withdrawal and the notification path to sites.
Safety
Draft
RFC-026
Translation workflow for patient education
Structure exists; this proposes the review process for translated clinical content.
Access
Draft

Datasets & registries

Cohorts are described publicly and queried where they live. No dataset listed here is downloadable from us, because no dataset listed here is held by us — the description is the shared artefact, not the data.

Example cohort descriptions showing the metadata each one carries. Names, site counts and scales are illustrative; no cohort below is currently enrolled.

GLIO-RETRO-4 Glioma 4 sites

Retrospective preoperative MRI with molecular annotation, described in OMOP so the same study definition runs at every participating centre.

DICOM OMOP CDM Federated query

Replication tracker

The uncomfortable table. Every headline result is re-run at a site that had no part in producing it, and the outcome is published either way — including when performance drops.

The rows below are worked examples, deliberately including a weakened and a failed outcome to show that those get published too. The deltas are not measurements, and sites stay unattributed until a real result exists and that site has agreed to be named.

Claim
Re-run at
Δ vs origin
Verdict
IDH status estimation from routine preoperative MRI
Independent site A
−0.06 AUC
Held
TME feature stability across scanner vendors
Independent site D
−0.19 AUC
Failed

Open questions

The things we don't know, stated plainly. Patients ask these constantly and the honest answer is that the field cannot yet answer them — which is exactly why they are the research agenda.

Do imaging-derived tumour features survive a change of scanner?

AURORA-GLIO

Our own replication says: often not. A feature that works only on the scanner it was developed on is a publication, not a tool.

See the replication result →

What outcome measures actually matter to patients over a lifetime?

Governance

Most registries record what is easy to record. Patient advocates on the council have consistently said the recorded endpoints are not the ones they live with.

Governance & ethics →

How to cite AURORA

Cite the module and version you actually used, not the project in general — module behaviour changes between versions, and a citation that can't be pinned to one isn't reproducible.

@software{aurora_glio_2026,
  title = {AURORA-GLIO: computational research infrastructure for glioma},
  author = {{AURORA Collective}},
  year = {2026},
  version = {0.1.0-pilot},
  note = {Research use only. Not a certified medical device.},
  url = {https://github.com/AURORA-NEURO}
}

Work with us on this

If you have a cohort, a method, or a disagreement with something above, all three are welcome. Disagreements especially — they are the fastest route to a correct result.

research@aurora.health
rfc@aurora.health
Cookies · federation-first by default

We keep essential cookies so the site works. Everything else — analytics on which guides help, and remembering where you left off in a guide — is off until you say yes. We never sell data, and nothing you read here is tied to a clinical record.