Between the world that produces evidence, and the society that acts on it, the substrate is the part that holds.

Every signed transition eventually traces to a wet lab, a clinical site, an instrument, a biobank, or a sensor in the field. The substrate is not a replacement for any of these. It is the layer that lets them inherit from each other instead of repeating each other's mistakes. A failed assay in Boston should weaken a dependent claim in Singapore by the end of the week — not by the end of the publication cycle.

• Wet labs Where syntheses, assays, and perturbations run.
• Clinical sites Where trials enroll and outcomes are observed.
• Instruments and sensors What produces structured measurements.
• Biobanks and cohorts Long-lived sources of patient and population data.
• Surveillance feeds Wastewater, wildlife, agricultural, environmental.

A protocol-execution rig hums in a synthesis hall at 04:00. Its readout writes back into shared state before the technician's coffee finishes.

Vela is the smallest set of rules that makes the rest possible: how findings are typed, how events are signed, how reducers materialize state byte-identically across implementations, how federation crosses jurisdictions, how AI-generated transitions get their own attestation tier. Three reducers — Rust, Python, TypeScript — replay the same cascade byte-for-byte on every commit. Dual Apache-2.0 / MIT.

• Keel The primitive type catalog (vf_*, vfr_*, vbr_*, vat_*, vco_*, vpr_*, vea_*, vcnd_*, …).
• Reducer Materializes the canonical-event log into byte-identical state across implementations.
• Signing rules Ed25519 identity, signature thresholds, revocation, key rotation, recovery.
• Federation grammar Peer-hub manifest, cross-frontier references, registry governance.
• Agent-attestation tier First-class signer tier for AI-generated transitions, with rate limits and audit.
• Conformance suite Test vectors every reducer implementation must pass.

Test vector vt_5076 replays in 11ms across all three reducers, identically, on every push to main. The protocol is the part that does not bend.

The Discovery Engine is the operational platform a researcher, agent, or reviewer sits at. It compiles artifacts into proposed state transitions; presents the reviewer queue, the contradiction catalog, the proof-of-replay; tracks predictions to their resolutions; calibrates agents against their reliability records. This is where the loop meets a person — the part of the system you can use today.

• Atlas: composition Frontier rollups, bridges, FAIR metadata; the materialized view over the substrate.
• Frontier cockpit Reviewer queue, source debt, trace proposals, benchmark evidence, proof commands.
• Diff packs Bundled change-sets; reviewers act on packs, not individual events.
• Trails Trajectories from question to outcome: data, tools, models, decisions, reviews.
• Agent attestations Envelopes pinning model, prompt, tool calls, and outputs that produced an artifact.
• FrontierBench Reviewer calibration, frontier predictive value, agent reliability. v0.1 deferred.

On the Early AD corridor, 5,531 findings, 51,787 evidence atoms, 26,606 typed DAG links, 300 contested findings, 500 gaps. The engine is a working object on one disease.

The terafactory is the physical body — Tesla-scale capital allocated to scientific infrastructure that reads from and writes back to public state. A facility compiler turns accepted protocols into plate maps, reagent orders, robot schedules, calibration lookups. Synthesis halls run continuously against shared frontier state. Pathogen surveillance feeds wastewater, wildlife, and hospital signals into the same surface a vaccine design pipeline reads against.

• Facility compiler Accepted protocols compile to plate maps, reagent orders, robot schedules, calibration lookups.
• Synthesis halls Robotic synthesis lines; biology floors; materials lines.
• Biomanufacturing Cold-chain manufacturing, GMP wings, candidate-construct production.
• Pathogen surveillance Wastewater, wildlife, hospital, agricultural feeds; rapid-response primer synthesis.
• Calibration registry Lot lineage, calibration log, synthesis history, federated, locally owned.

A failed run in a synthesis bay writes back the same shape as a successful one. The factory does not lose its negative results to a private log.

Foundations underwrite trials with deposit-conditional clauses. Regulators inspect state histories alongside trial data. AI labs train against attested corpora and sign their own deposits. Patient foundations make body-clause grants. Journals cite canonical state, not just papers. Sovereign-science programs federate into the same surface their hospitals already read from. The substrate does not abolish gatekeepers. It changes what is being kept.

• Foundations Conditional-deposit clauses, milestone funding tied to signed deposits.
• Regulators State histories as inspectable support for IND, CMC, DSMB, IRB, IACUC packets.
• AI labs Foundation-model training corpora with attestation; agent-attestation deposits.
• Patient foundations Body-clause grant programs requiring writeback at the gigafactory boundary.
• Sovereign programs National AI-scientist initiatives federating into the public substrate.

On a Tuesday in 2034, a foundation officer signs an amendment that releases a tranche against three signed deposits. The amendment is in the same surface the clinician reading the new recommendation is using.