Token routing architecture

Route intent without hiding asset, policy, signing, or settlement risk.

Build token router nodes as transparent state machines for asset identity, discovery, quotes, simulation, policy, execution, finality, and reconciliation.

Token routing is not one protocol. A router may compare venues, select chain endpoints, coordinate bridge messages, build transactions, or invoke a custody system. Architecture begins by naming the object being routed and the stage the router owns. A proposed path is not settlement, and a successful API response is not final value transfer.

Use canonical asset identifiers that include chain, contract or native asset, standard, and decimals. Treat wrapped and bridged representations as explicit relationships rather than automatic equivalence. Registry changes are high-impact and need provenance, versioning, approval, and rollback.

Keep discovery, quoting, simulation, policy, execution, settlement monitoring, and reconciliation separate. Bind every stage to an intent ID and immutable versions of the quote, registry, policy, and transaction payload. If the quote expires or state changes, return to the relevant stage instead of silently executing a different route.

Signing should remain outside a general pathfinding process. Prefer unsigned payloads that a user or constrained signer approves. Show exact contracts, approvals, recipients, amounts, deadlines, and chain context. For cross-chain work, model source and destination finality as explicit states with evidence and idempotent retry behavior.

A router is an evidence pipeline

For every intent, preserve the exact asset identities, candidate paths, quote normalization, policy decision, payload, signer, submissions, confirmations, and reconciliation result.

01Canonical asset registry
02Expiring normalized quotes
03Deterministic route policy
04Payload-derived signing summary
05Protocol-specific finality states
06Idempotent retry and reconciliation
Operating questions

Make the hidden assumptions visible.

Each answer should map to a test, metric, policy, or runbook—not merely to a design claim.