Radix Week in Review: Rehearsing the Restart

The week in brief

• The fix for the bug that stopped Radix is written, tested and public, as a change to the Radix Engine named Eagle Ray.
• It gets rehearsed on Stokenet, Radix's test network, over the coming days. That is the first dated step towards a restart.
• The Radix Accountability Council has confirmed the cause and says the remaining work is delivering the fix safely and getting the validator operators to adopt it.
• A written account of what happened will follow the restart, and why the audits missed the bug is being looked into separately.
• The DAO's incorporation went ahead anyway. RADIX DAO LLC gets filed in the Marshall Islands on Monday.
• Every number in this issue was read from the stopped ledger, which turns out to answer any question that names a moment in its history.

Radix mainnet, the live network that holds real money, has been stopped since the evening of 31 August. Someone found a bug in the Radix Engine, the software that runs every transaction on Radix, and used it to take tokens out of accounts they did not own. In under an hour they emptied the network of its bridged assets, which are wrapped copies of tokens from other blockchains, and sent them out through a bridge called Hyperlane. The operators who run Radix's validator machines shut them down that evening. The drain and halt page has it transaction by transaction.

This week was about the repair, and it has a shape now that it did not have on Monday.

The fix, and the rehearsal

Restarting needs four things in order: someone writes a fix; it goes into a named protocol update; someone releases a new version of the node software containing it; and enough validator operators install that version and signal they are ready, which is how a Radix upgrade activates.

The first two are done, in one pull request. A contributor called 0xOmarA opened pull request #2093 on the Radix Engine's repository on 2 September, adding a protocol update named Eagle Ray. Radix names its upgrades alphabetically, after Anemone, Bottlenose, Cuttlefish and Dugong. The change is small: before running a method call, the engine now checks whether the caller is allowed to see the thing it is calling the method on, and refuses if it is not.

Step three is where it sits. Nobody has merged the pull request, and the newest release of babylon-node, the software the validator operators actually run, is still v1.3.0.5 from June.

What moved this week is the rehearsal. Daffy, a community contributor who maintains the DAO's governance repository, told the Radix Developers group this morning to expect Stokenet, Radix's public test network, to go unstable and go offline over the coming days while they run the upgrade through it. Stokenet is working normally, so breaking it on purpose is the point. It is the first thing anyone has put a timeframe on since the network stopped.

What the council has said

The Radix Accountability Council, the volunteer group currently running the project, gave the fullest account on 3 September. It has identified and confirmed the cause, a flaw in how the Radix Engine hands out references to vaults, a vault being the container that holds tokens inside an account. The patch, it said, is the easy part. The fix has to reach the network without giving anyone a window to use the bug again first, and then enough validator operators have to read the plan, agree with it and install the software. About 100 independent operators run those machines, so restarting means 100 separate decisions.

Daffy added two things on 5 September. A written account of the bug will come after the restart, and why the paid security audits missed it is being looked into separately. That is the first firm commitment to publishing one.

Reading a stopped ledger

Almost everything that displays Radix data reads it from the Gateway API: wallets, explorers, price dashboards, this site's charts page. Ask it where the ledger is, the ledger being Radix's record of every account and what it holds, and the node answers from the last transactions it agreed. Ask what the ledger contains, and the answer comes from a separate database that copies the ledger as it goes, is allowed to be 12 minutes stale, and is now five days behind. It refuses.

The useful part is that the staleness rule is skipped for any request naming a specific point in the ledger's history. Ask for the validator list as of state version 557,840,622, the last state the network reached, and the same endpoint answers in full. No account or key is needed, and every number below comes from that request. It also means anyone can check what their own account held when the network stopped, which is the question most people have been asking all week.

One thing to fix at home first. A program asking both questions gets a timestamp that looks current attached to a list of nothing, and if it adds that empty list up, the error becomes a zero. This site's charts page did that for six days, reporting no validators and no XRD staked as though someone had measured it. The code changed this morning: it now raises an error instead of returning a half-empty list, the charts page says the data is unavailable, and the weekly series has the bad entry removed.

The DAO's paperwork went ahead

The company that built Radix is handing it over to a DAO owned by the people who hold XRD, Radix's own token, and part of that is registering a legal entity for the DAO to be. On 5 September a council member said the contract with MIDAO, the firm handling it, is signed and the fee is paid, and that filing RADIX DAO LLC in the Marshall Islands starts on Monday. The registry takes four to six weeks, so the certificate should arrive in the first half of October. None of it needed the network.

The other half of the handover does. Ratifying the DAO's governance rules needs a vote by XRD holders, and that vote runs on the network, so it waits.

Hyperscale's author turned down the payment

Hyperscale is a replacement for the part of Radix that puts transactions in order and agrees them, written almost entirely by one developer who goes by flightofthefox. In August the community voted to pay him in five milestone instalments. On 3 September, days before the first was due, he said he will not take the money. I just want to work on the tech and try to keep my sanity and dignity, he wrote, and said he will build the rest anyway and the DAO should spend the tokens on something else.

The group argued about it all day, mostly over whether Hyperscale is worth finishing if the network it replaces does not come back. His commit history shows a pause rather than an answer: 338 commits to the virtual machine in the week to 30 August, 6 in the week since.

The week on the ledger

The last numbers Radix produced, against 29 August, which is two and a half days earlier rather than a full week. They will be the same next week unless it restarts first.

A validator takes a fee out of the staking rewards it passes back to the people who stake with it, and announces a change in advance, naming the epoch at which the new rate starts. An epoch is a stretch of about five minutes, and epochs only advance while the network runs. Three announcements are queued: Leaf Node going from 1% to 100%, Apollo Pool from 20% to 100%, and Avaunt Staking from 2% to 25% on 139 million XRD. All three take effect shortly after the restart, and nobody staked with them can move out beforehand.

Concentration watch

Eight validators between them held a third of all staked XRD when the network stopped, the same eight as in each of the three previous weeks, and 15 held half.

The ledger

This series records dated predictions and scores them. One came due today, and it is a hit. On 18 August it predicted that StakeSafe would raise the fees on all three of its validators as announced, and it did: Amsterdam and Rotterdam are charging 25%, Seed Node 100%, landing about a day before the network stopped.

Two new predictions, both due 20 September: that mainnet processes transactions again, and that a release of babylon-node containing Eagle Ray is published. The second has to happen before the first can. The scores and the exact checks are on the series index.

This week on the wiki

Hyperlane Asset Drain and Network Halt (August 2026) has been rewritten every few hours since it happened and now runs day by day. The part to read is the section on checking your own account: the Radix dashboard cannot tell you what you hold, because it asks the Gateway the question the Gateway is refusing, but two requests naming the last state version can.

What to watch

• Stokenet going offline on purpose, this week, while they rehearse the upgrade. Confirmed by Stokenet reporting a protocol version later than Cuttlefish.
• A babylon-node release containing Eagle Ray, which has to come before mainnet restarts. Confirmed on the releases page.
• The Marshall Islands certificate for RADIX DAO LLC, in the first half of October.

Corrections

Last week's issue said the StakeSafe prediction could not be settled until the network came back. It could: a request naming the last state version settles it, and it is scored above.

Sunday's snapshot script recorded zero validators for the week, because the Gateway returned an error and the code counted it as an empty result. Removed, and the script now refuses a reading with no validators in it.

Next week has one shape to it. A release of babylon-node carrying Eagle Ray, a rehearsal on Stokenet, and 100 operators deciding to start their machines again.

Radix Week in Review series, Issue #11: ← Previous: Issue #10, Aug 24–30, 2026 · All recaps · Subscribe

Previously: A rebuilt Stokenet caught up on three protocol updates in under two hours, because what gates an update is a stake threshold rather than a date.