The Week the Chain Froze, and the Day It Came Back

All times in this article are UTC, the same clock the engineering thread and the release pages use. Every number is a measurement from our own production nodes or from the public record, and each one is timestamped in the text.

What actually froze

BTX validates its new proof of work in two layers. Machines with a qualified GPU replay each block's MatMul computation themselves. Everyone else, which means most nodes, follows a signed frontier: an attestation signed for each block by the network's attestor, then fanned out by archive nodes that store and serve those confirmations.

That design has one failure surface, and August found it. When attestations stop flowing, every node without a GPU freezes in place while looking perfectly healthy. This can happen because the attestor pauses, because too few archives serve, or because peers that could serve are never asked. Blocks may even keep being mined above the frozen frontier, but without signatures they are islands. In the final stall, one operator publicly reported mining six blocks above the frontier. All six were later lost as islands, real work spent on a chain that could not become canonical.

Here is what the last freeze looked like from our production explorer:

• The signed frontier stopped at height 191,690 at 10:58 on August 17.
• Headers kept arriving above it. Our node saw peers claiming up to 191,716, proof that mining continued while confirmation supply did not.
• Transactions sat unconfirmed for hours in mempools across the network. They were never at risk. They were simply waiting for a chain that could accept them.

This was the deepest point of a stall era that had been building for a week, documented block by block in the public engineering thread on the BTX repository (pull request #105, which gathered 133 comments of operator reports, ours among them).

What v0.33.3 changed

At 14:46 the core team tagged v0.33.3, and sealed binaries followed within the hour. Three changes matter most, in plain terms.

1. A hard difficulty floor at block 191,714. The first block after the frozen region must meet a fixed target roughly three times harder than the blocks before it (floor difficulty of about 1.49e-6 against roughly 5.5e-7 prior). From 191,715 onward, difficulty adjusts on a long half-life. A stalled network can never again unwind its difficulty toward zero and reopen the lottery that made cheap islands possible.
2. One signed hash per height. The attestor now signs exactly one block per height. This ends the "twin lottery," where competing blocks at the same height split the network's attention.
3. The stall paths are closed. Nodes without a GPU now fetch and follow attested catch-up correctly, archives keep serving during replay, and the retry loops that could starve a node from the inside are gone.

An upgrade deadline came with it. Any node still below v0.33.3 cannot accept the floor block. It pauses at 191,713, paused, not broken, until updated.

The recovery, hour by hour

All measurements are from our own nodes' logs and RPC on August 17. Nothing below is estimated.

Two details from the crossing are worth recording. First, catch-up over the frozen range ran at body-download speed. Our node connected 23 banked blocks in 80 seconds the moment the new engine accepted them, because archive peers had been holding the bodies all along. The freeze was never about missing data. It was about missing signatures. Second, the first confirmations of the new era rode inside the floor block itself. The hardest block of the day was also the one that ended the wait.

The network that evening, in numbers

Measured at 20:16 from our explorer's node, a single vantage point, together with the easybtx.com node census, which rechecks the network every 30 minutes. The true network is larger than any one node can see.

• About 60 reachable nodes: 54 community nodes discovered over the BTX network itself and verified with a real handshake, plus 8 curated directory nodes.
• More than 10 countries, led by the United States (13), France (9), Germany (7), and the Netherlands (7).
• 47 peers connected to our node, and 25 of them, meaning 53 percent, already ran v0.33.3 less than six hours after release.
• 11 attestation archives visible, with 9 actively feeding us signed confirmations. Hours earlier, that had been one of the scarcest numbers on the network.
• Our own node served confirmations to 35 peers and has uploaded 3.4 GB of chain data to the network so far.

Our part of the fleet crossed the same day: the explorer at btxscan.io, which now verifies its own freshness against the signer quorum directly instead of against other websites; BTX Node 0.6.10 for Mac with the v0.33.3 engine built in; and the BTX Keeper, a one-command installer that now sets up a serving node in seconds with no developer tools required. Windows and Linux node builds follow on their own trains and will pause at 191,713 until then.

What opens up next

A network that just proved it can heal is worth building on in the open. Here is the direction, without promising dates.

A public home for the code. The applications around BTX, meaning the explorer stack, the node app, and the Keeper, are moving to a dedicated open-source organization, published as clean releases from day zero. The Keeper goes first: it is small, self-contained, and it is the piece whose adoption most directly strengthens the network. The explorer and the node app follow.

Mining, opened in stages. The consensus engine that actually validates and mines, btxd, is already public upstream, and every sealed release we ship is verifiable against it. What we add around it, and intend to open, is the application layer: the miner app's staging and provenance tooling, which is how a release proves it contains exactly the engine it claims, plus the pool-mode plumbing and the run-your-own recipes. The security-critical core goes public last and loudest, together with an independent human audit by a senior engineer who already has access, so the code and its review arrive as one artifact. Open source is how this stack earns trust faster than any claim could. Sequencing is how we do it without shipping our own mistakes as attack surface.

Making it easy to give a little. The recovery was carried by ordinary machines serving signed confirmations. That lesson becomes product: the Keeper today, a trusted node directory with operator login next, and builder tooling after that, including SDKs and assistant skills that can set up a node end to end. One principle stays fixed: recognition, not rewards. There is no payment and no token for running any of this, and anyone who tells you otherwise is not speaking for this project.

Where things stand right now

A closing snapshot, taken at 22:05 on August 17, about four and a half hours after the floor block.

• Chain height 191,979, which is 265 blocks since the floor, arriving at roughly a block a minute. The mempool is empty.
• 60 percent of the peers our node sees now run v0.33.3 (27 of 45), up from 53 percent two hours earlier. The upgrade wave is moving on its own.
• Network proof rate has climbed from about 19 to 98 MatMul proofs per second since the floor, a five-fold rise in three hours. Miners are coming back. These are small numbers with heavyweight units, since every proof is a real GPU computation.
• 10 archives visible, 7 feeding us, and our node serving confirmations to 34 peers.
• The community census counts 55 active nodes tonight, the highest reading of the day.

The chain that spent a week frozen now produces a block a minute, confirms transactions within minutes of receiving them, and is watching its own mining power return by the hour. That is the state of the network, measured, timestamped, and holding.

The honest footnote

Today the frontier is signed by one attestor. That is Phase 1 of the network's own published path, and later phases add attestors and raise the required quorum. A research page that hid this would not be worth reading. We publish what we measure, including the parts still maturing, because a network that recovers in an afternoon has earned precise reporting, not marketing.

Sources: our production nodes' RPC and logs (the btxscan.io explorer node and the Apple-silicon archive node whose field measurements are reported upstream); the public engineering thread, BTX repository pull request #105; and the v0.33.3 release notes. All times UTC; measurements timestamped in the text.