Build the risk engine before the interface

Protocol invariants are the real specification

Write the conditions that must remain true after every instruction: conservation rules, collateral requirements, share accounting, authority checks, fee limits, and pause behavior. Each invariant needs normal, adversarial, and recovery tests. Interface requirements come after this ledger of allowed state transitions.

• Define rounding direction and acceptable dust for every asset calculation.
• State who can change parameters, upgrade programs, or pause operations.
• Test repeated, reordered, partially failed, and duplicated transaction paths.

Oracle policy needs explicit failure states

Price integration is more than reading an account. Define accepted publishers, staleness limits, confidence intervals, fallback behavior, and the action taken when a feed is missing or disputed. Liquidations should not proceed on an undefined price policy.

Liquidation design joins economics and infrastructure

A liquidation can be economically valid and still fail operationally. Model account contention, priority fees, keeper incentives, transaction simulation, confirmation, and retry rules during volatile periods. Observe both protocol state and transaction delivery so the team can tell market risk from infrastructure failure.

RPC and indexing are part of the product boundary

Wallet balances, positions, history, quotes, and alerts depend on reads that may arrive late or disagree across providers. Use more than one endpoint where the business impact warrants it, keep an index rebuild procedure, and record the slot or commitment behind user-facing state. RPC Edge is a separate Builderz venture focused on this operating layer.

• Separate transaction submission from read-heavy indexing workloads.
• Track simulation errors, blockhash expiry, confirmation time, and provider divergence.
• Design a degraded mode before a provider incident happens.

Security review begins before an external audit

External review cannot replace threat modeling and local evidence. Keep instruction-level authorization tests, property tests for accounting rules, upgrade and key-management procedures, dependency ownership, and a release checklist. Give reviewers a small system with written invariants rather than a moving target.