The substrate-level rules AURORA refuses to negotiate. Equity audits run inside the training loop. Patient-advocate seats are permanent. Releases that fail the equity gate are blocked, not warned.
Stratified performance is part of the loss function. Releases that fail equity thresholds are blocked.
Every recommendation explains itself; every override is logged with reason.
Two permanent patient-advocate council seats with veto rights at release.
Cohort policy travels with the data. Subsystems refuse to run on records without the right consent.
Data stays at the institution. Compute travels to data, signed and audited.
MIT at public alpha. Model cards, equity audits and replication reports ship with the code.
Stratified performance is part of the loss surface, not a post-hoc audit. If a model can only reach its top-line metric by performing badly on a subgroup, training does not converge.
For every released module we define a stratification set — typically ancestry, sex, age band, socioeconomic proxy, and access proxy. During training, the loss function carries a penalty for predictive disparity across those stratifications. At release, the stratified performance is reported per stratum (not just aggregate), and a drop below the equity threshold defined in the module's RFC blocks the release.
There is no exception path for marketing reasons. Thresholds are revisable — but only by RFC, in public.
If we cannot reach the top-line metric without hurting a subgroup, the top-line metric is wrong. Not the subgroup.
Consent metadata is a first-class object in the AURORA substrate. Every record carries the cohort policy it was collected under. Subsystems with an explicit consent dependency — equity stratification, organoid sync, identifiable surgical video — refuse to run on records that lack the matching policy.
Federated training honours the same constraint. If a cohort's policy disallows a particular use, the federation runtime will not include those records in that training run — and the audit log says so.
Stratified performance per stratum. If any drops below threshold, the release is blocked. No exceptions.
Every training record carries a policy. If any record's policy disallows the use, the run is invalid.
Patient-facing surfaces require advocate-seat sign-off. Plain language. No exceptions, no rush.
Ethics-board complaints reach independent ethics-seat holders who are not on the core team. Patient-advocate-desk complaints reach the two permanent council seats. Privacy concerns reach security@ and are triaged within 72 hours.