hyperscale-rs, the community-built Rust implementation of a sharded Radix network, splits a shard in two when it grows too large, and merges two when they shrink. A change its lead developer, flightofthefox, merged today decides what happens to transactions caught across a split or merge. Until now a shard kept taking cross-shard transactions right up to the moment it split, and the ones caught mid-way were sorted out afterwards. Now the shard drains first: it stops admitting transactions to and from other shards, waits until every transaction it has already agreed to has its answer from the other shards, and only then splits. Transactions that touch that shard alone keep running throughout. With nothing at fault, the drain takes at most 444 seconds, about 7½ minutes. If the splitting shard is waiting on a shard that has halted, and the drain misses its deadline, the split is cancelled rather than forced, so no transaction can commit on one shard and abort on the other. The design, including what the pause costs: github.com/hyperscalers/hype…

On Radix, who may call a method is written once in the blueprint: one block maps each method to the roles allowed to call it. The engine checks the caller's badges against that rule before the method runs. A Solana program has to confirm that every account it is handed is owned by the right program and signed by the right key. Anchor makes those checks when the developer picks the right type for each account. radix.wiki/contents/tech/com…

Scrypto developers now have a free pre-audit kit: scrypto-audit-kit, open-sourced this week by bigdev, who also built the Radix Guild task marketplace. Scrypto is Radix's Rust SDK for smart contracts. It reads a blueprint in two passes. The first is a fixed set of rules that needs no API key and flags patterns such as a component published with no owner, roles that can rotate themselves, floating-point maths and hardcoded addresses. The second, optional, has Claude work through an 11-class checklist: authorisation, reentrancy, decimals, resource handling, oracles, slippage and more. Both feed one report that a person or an AI agent can read, and an MCP server lets coding agents run the audit, fix what it finds and check again. A report can also be turned into a Radix transaction that records it on ledger, tied to a hash of the exact source it read. It is alpha, and the README says it does not replace a human audit: it is a cheap first read before paying for one. Install with pip install scrypto-audit-kit (Apache-2.0): github.com/bigdevxrd/scrypto…