Don't trust the number. Re-run it.

Data driven keeps the result. Compute driven ships the derivation.

The name is an argument, and it is not that data-driven work is irreproducible. It often is reproducible — when the inputs, the transformations, the environment and the method have all been preserved with enough discipline. The difference is where that discipline lives: alongside the artifact, or inside it.

So, stated as a test rather than a mood:

And the limit, because this is the sort of page that has to state it: re-derivation proves fidelity, not correctness. A wrong computation re-runs perfectly and is still wrong. Deriving the same identity tells you that you got what we got — it does not tell you that either of us should have wanted it. What it removes is the argument about what happened, which is the argument that usually eats the room.

A road, and a window. All the way down.

This is a WRL world — the language the stack writes meaning in. Square brackets are places. Arrows are roads between them. It is a map, and the map is executable.

Now look at the shell in the picture at the top of this page: windows standing as signs along a road you drive through. That is the same drawing. Not a metaphor borrowed for the marketing — the language, the runtime, the evidence and the interface are all the same two things.

The shell's own documentation puts it better than we can: “One likes going; the other likes being gone to.” Neither half is worth much alone — a hash without a film is a claim, and a film without a hash is a story.

Six rules. We break them in public or not at all.

These are not aspirations. Each one is here because it changed something we shipped — usually by making a page say something less flattering than the draft did.

1. Re-derive, don't trust. If a claim cannot be computed again by someone who does not work here, it is a rumour with good typography.
2. Say what is not built. Every status on every page is measured, sourced, or marked not built. The absent things are listed beside the present ones, at the size they actually are.
3. Subtraction is the feature. 364 MB is not an optimisation. It is 200 MB of test suites, every debug package and every 32-bit library, decided against.
4. A claim carries its date and its machine. “It works” is not a result. “Measured on a pristine image, 10 August 2026, and here is the string it printed” is one.
5. Quote the unflattering half. The verifier costs 14 MB. The interpreter it needs costs another 250. A page that mentions only the first number is lying with a true fact.
6. Never fake a success. A control that reports something it did not do is worse than one that plainly fails. We deleted our own before writing this down.

Two images. Pick the one you need.

Three ways in. One of them is proven.

Every T&R boot on record is a virtual machine. Writing the image to real hardware should work — it is a GPT image with a UEFI partition — but nobody has done it, so those two routes are marked untested rather than supported.

Subtraction, done once, at build time.

FreeBSD's own cloud image enables firstboot_pkg_upgrade, so every instance re-downloads its base packages the first time it starts — forever. We patch at build time instead, and leave out the ~200 MB of test suites, every debug package and every 32-bit library a server guest will never open.

Six parts. Not all of them are software.

T&R is the distribution and the shell; the five beneath it are the reason the distribution exists. They are not the same kind of thing, so each card says what role it plays and where it actually is — because "part of the stack" and "in the image you just downloaded" are different claims. ComputeDriven is not a seventh box. It is the discipline the six are held to.

The parts we have not earned yet.

A download page that only lists wins is asking you to install on faith. These are the holes at the size they actually are.