Most software hands you a conclusion and asks you to believe it — a score, a
benchmark, a dashboard. We build the other kind:
systems whose claims you can derive again yourself, on your machine, and
get the same answer or a different one.
What you end up with is a world — your machine, its state, and the record
of how it got that way. ComputeDriven is one destination for that world and two
ways to reach it.
Two entry points lead to one destination. Both are a download. T&R, the operating system, is downloaded and booted. Super (CD), the compute surface, is a desktop application: downloaded and installed alongside whatever you already run. Both lead to [World] Cloud, where a world is hosted, forked and restored on any machine.
Two entry points. One destination. Both are a download — only one of them replaces what you boot.
A world is not a disk image. It is the machine, its state, and the record of how it got that way — and because that record is a graph rather than a snapshot, it can be moved, forked and replayed rather than merely copied. Cloud is where that graph lives when it is not on your desk.
One world, any machine — hosted, forked and restored without provisioning.
No instance to size, no region to pick, no disk to attach. You hand Cloud a world and get an address back; you hand it an address and get the world back, on whatever hardware is in front of you. The thing being stored is small — the semantic layer of a real world measures in hundreds of kilobytes, not gigabytes — which is why this is priced as a place to put something rather than as a bucket.
They are not tiers and neither is a trial of the other. They are two honest answers to a different question: do you want the world on metal you own, or do you want to drive it without changing what you boot?
FreeBSD 15, assembled from pkgbase rather than installed. No installer, no first-boot download.
The 626 MB image carries its own verifier, which is the half that lets the sentence at the top of this page be taken literally: you can re-derive what it claims without asking us. This is the door for people who want the machine, not a login.
A desktop application — the compute surface, holding control, placement, evidence and authority above interchangeable engines.
Super sits over whatever is actually doing the work and keeps the parts that outlive it: what was decided, what was permitted, what it cost, and what the evidence was. It reaches the same world Cloud holds, from a machine you have not reformatted.
super.computedriven.com
A claim is compute driven when the artifact carries enough — canonical input, executable semantics, a derived identity and its provenance — for an independent machine to derive the claim again.
Re-derivation proves fidelity, not correctness. A wrong computation re-runs perfectly and is still wrong. What it buys is narrower and more useful than truth: it removes us from the chain. You do not have to trust that we ran it — you can run it.
It is also what makes an agent a different kind of thing here. Most AI products hand a model some tools and call the bundle an agent — which names the parts and not the property. In a world that keeps its derivation, an agent is a continuing position in that world: a state you can establish rather than take its word for, the scoped authority associated with that established state, and a record connecting what it changed to what permitted the change. Change the model underneath and the definition stops deciding the identity question for you — it hands you something to test instead. Did the successor re-establish the state, the authority, the obligations and the continuity required to carry on? Model + harness makes swapping the model an identity change by fiat. This makes it an empirical question, which is the part the usual definition cannot express at all — and one we have not yet answered. The invariants a position is established under →
The long version of this argument, with the ladder, the certificate and a list of the parts we have not earned yet, lives on the T&R site — because that is the product that has to survive it. Read it there →
Cloud, T&R and Super are what you buy or boot. These are what they are made of — published separately because a component that cannot be inspected on its own is a claim, not a component. The stack has converged on three verbs, each owned by one part: WRL describes understanding. TRVM reduces and derives it. Graphonomous stores, connects, explores, tests and evolves it. ComputeDriven is the world the three run in, and Super is the environment around them that lets humans and artificial beings actually use them.
Everything else — the specs, the docs atlas, the individual protocols — is in the portfolio menu at the top of this page, or at docs.ampersandboxdesign.com.
Five pieces on how this is built. They are arguments rather than announcements — each one is the reasoning behind a decision the products above are already living with.
Their sites are linked above and both products carry an in dev status for a reason — what is behind those doors is being built, not finished. T&R is the one you can have today: it is downloadable now and owes you nothing, no account, nothing to sign up for.
For the other two, this is the list — a person reading a form, not a drip sequence. Tell us which one and what you would use it for; that is what decides the order they get finished in.
That is the one gap we cannot close ourselves — every measurement we
publish was taken on our machines. Boot reports are filed as issues
on the distribution's own repository, with a template that asks what
trvs doctor
printed.
Third parties on this page, in full.
The form above posts to formspree.io; nothing else here
calls out at load, and the page renders fine with that host blocked —
only the form stops. Separately, the host injects
static.cloudflareinsights.com at the edge. It is not in
this file and not in what the server returns to
curl, and we cannot switch it off from this repository,
so it is named here rather than denied.