Jun 2026
The re-run is the proof
Refuse to keep a ledger and three things you'd normally build a company for fall out for free.
Bitcoin burns around 150 terawatt-hours a year. That is more electricity than the Netherlands, spent on one task: getting strangers who distrust each other to agree on a single list of who owns what. The hashing is not useful work. The burn is the cost of agreement, and that is the whole game.
So we asked a dumb question. What if you just don't share a list?
The cut
There's a theorem for this, called CALM, proved by Hellerstein and Alvaro. It says a distributed system needs zero coordination as long as it only ever learns more and never takes anything back. Look at what a network actually does and almost all of it is exactly that. Content you've seen stays seen. Names pile up. Reputation only grows. The heat of popular data climbs and fades on its own. None of it needs a vote.
Two things break the pattern. A coin can't be spent twice. A unique name can't be claimed twice. Those, and only those, get pinned down in one small bonded service. And even that isn't a ledger. It's a blind-signed token (the same trick David Chaum shipped in 1982), a set you check membership against, and a one-line reserve check. No balances, no history of who paid whom. The entire state of the money is a set and two counters.
Everything else stays where it lives and gets checked on the spot.
The trick
Here's how you trust a computer you don't own. You make it run a deterministic program, then you run it again yourself. WebAssembly pins this down at the spec level: no fused multiply-add, no 80-bit float drift, NaNs canonicalized, so the same input gives the same bytes on every honest machine, forever. "Did this run right" stops being an argument you have to win and becomes a fact one honest re-run settles. A liar is caught by a single signed receipt. No race to be first, no majority to bribe, no reorg. Proof-of-work does fake work to buy an ordering nobody asked for. We do the real work, and the re-run is the proof.
Then it pays you twice more
This is the part that got us. The same determinism we built for trust is quietly two other things.
A time-travel debugger, free. Mozilla spent years building rr, a tool that records a program so you can replay it exactly and step backwards through a crash. If every run already reproduces bit for bit, you skip the recording. The receipt is the repro. A bug report becomes a hash anyone can re-run and watch fail.
An LLM senate, free. Pin a model's weights by their hash, fix the seed, set the temperature to zero, and the model's output is itself a deterministic program. Replayable. Hashable. So a council of models could read the network's signed records and decide things, and anyone could check exactly what it decided by re-running it. Honesty about what's wired today: the council that actually allocates our grants is deterministic arithmetic — a karma-weighted quadratic-funding fold, replayable for exactly the same reason — and no model sits in that loop yet. The senate is the same trick pointed at bigger programs. You might hate the model's taste. You can't be lied to about its verdict.
One mechanism, determinism plus content-addressing, and you've paid for trust, debugging, and machine governance at once. The chain spent a continent of power to avoid trusting anyone. We spend a re-run, and pocket two superpowers on the way out.
The catch
A re-run only settles things if someone actually re-runs. We sample, cheap when a provider has a track record, certain when it doesn't, backed by a bond so cheating loses money on average. The reserve behind the money is checkable, but the reserve itself is the one spot you trust an edge. The council's determinism is proven. Its taste isn't. That line between math and economics sits exactly where the theorem says, and not an inch further.
Verse is the CALM theorem turned into an operating system. Everything that can be left alone, is. The two things that can't, aren't. And the re-run is the proof.