Architecture
A layered system: a local-first state engine at the edge, a reconciliation service that merges node states into a canonical ledger, an institutional integration boundary, and an audit layer that observes every layer beneath it. Layers communicate through defined interfaces rather than shared internal state.
State model
Transaction state is append-only and content-addressed: each state transition references the transitions it depends on, so the causal order of events is recoverable independent of wall-clock time or network arrival order. This is what allows a node to operate correctly while disconnected.
Settlement orchestration
Settlement is coordinated by the Reconciliation Engine, which sequences merges deterministically and records the outcome of every settlement decision in the audit layer. Orchestration logic is separated from the state engine so settlement rules can evolve without changing how state itself is recorded.
SDK / API model
The Institutional API is being designed around a small set of primitives — submit, query, reconcile, observe — rather than a broad surface mirroring every internal capability. An SDK layer on top of that API is planned but not yet published; the API contract is being validated against integration scenarios first.
Observability
Reconciliation lag, divergence windows, node health, and audit-log integrity are exposed as structured signals rather than derived after the fact from logs. Observability is treated as part of the infrastructure contract, not an operational add-on.
Fault handling
Node failure, network partition, and reconnection are handled as expected operating states rather than exceptions requiring manual intervention. A node that fails mid-write recovers to its last durable state on restart; a partitioned node continues operating locally and reconciles on reconnection.
Security boundaries
The Institutional API is the only externally reachable boundary into the system; internal layers do not accept direct external input. Every write is signed at its point of origin, and every access to the audit layer is itself audited. Security disclosure follows the process on the Contact page.
Testing methodology
Reconciliation correctness is tested against generated divergence scenarios — partitioned nodes, out-of-order delivery, duplicate submission, conflicting concurrent writes — rather than only against the happy path. Financial-integrity behavior is treated as a release gate, not a post-hoc check.