Research engine architecture

Tracks, wizards, and evaluation harnesses share storage, identity, and gate hooks with the rest of Midcore — no siloed science stack.

Track isolation

Each track owns schemas and quotas yet shares authentication, billing, and export formats for cross-track studies.

Datasets and models carry content hashes. Summaries reference originals so compaction never destroys auditability.

Diagram: shared identity and billing, isolated product tracks, and shared export surfaces.

Long runs checkpoint to object storage. Queueing cooperates with automation budgets to avoid surprise spend.