The standard

Four kinds of number

Every figure on these pages carries a mark saying which kind of number it is. live is computed in your browser, this instant. exact is a closed form, or integer-exact arithmetic. measured is sampled from the run in front of you, and comes with its error. stored is a constant from a longer computation, verified against an independent reference before the page shipped. Nothing is typed in because it looks right.

Nothing ships unchecked

Before an instrument goes up, an independent reference — written separately, in a different language, by a different route — has to agree with it. Most instruments then carry a self-test you can run from a button on the page. A check that runs on your machine is worth more than a promise made on mine.

Build it, then try to break it

One agent builds. Others are then set against the build with a single instruction: find what is wrong. Not "review", not "confirm" — find what is wrong, and quote it. A reviewer's worth is measured by what it finds, so reviewers here are rewarded for finding and never for agreeing.

What that has caught, so far: a page that claimed every instrument carried a self-test, when one did not; a clip described as a seamless loop that, measured, was among the worst loopers in the set; a document about the importance of counting things whose own character counts were each off by two. Each of those would have shipped on a reviewer's nod.

Say what it cannot claim

Each project states, in the same place and the same shape, the claims it declines to make. The simulation is an intuition pump, never evidence, and certain words are banned from everything written about it. The map of clinical development never presents an absence as a finding: a view the data cannot support is drawn as withheld, not guessed. The network demo says on every page that all of it is simulated. The artworks say they are artworks, not figures. A refusal costs something — an emptier map, a quieter claim — and that cost is the point.

Say what you own

Where the work touches something I have money in, the page says so, in the same place every time and before the argument starts: the Varseta-M briefing carries "the author is long CTMX" on its masthead, in its refusals and in its caveat. A disclosure at the bottom is a disclosure timed to be read last. This one is timed to be read first, because a reader who knows my exposure can discount my conclusions, and that is the point of telling them.

Measure, don't guess

When the mark changed colour, its contrast was measured before and after. When a clip needed to loop, the seam was measured pixel by pixel across candidate lengths — and the piece whose own source code said "seamless" turned out to describe one sub-cycle, not the whole; it measured worst. When a still needed the whole of a circle inside its frame, the framing constant was found by rendering candidates and sampling the border, not by reading the shader. When a comment and a measurement disagree, the measurement wins.

Reproducible by construction

Every asset on this site can be regenerated from a script with its parameters recorded. Every clip carries a sidecar naming the exact command that remakes it, frame for frame. The simulation burns its seed and parameters into every frame it exports, so a clip in a feed is an experiment anyone can re-run. A number that cannot be re-derived is a rumour, however confidently it is stated.

Built with AI, checked by adversaries

The division of labour is specific. A person directs the work and decides what ships. Models build. Other models are set against what was built. Independent references arbitrate. The person reads the review as critically as the build — a reviewer that reports a false problem is a defect too, and there have been those.

Everything here is built in close collaboration with Claude models (Anthropic), across research, code and prose. That is a description of the tools used, not of a relationship.

Use it

Nothing above is proprietary. Take the vocabulary, the review shape, the rule about absences. And hold this site to it: if a number here fails its own standard, say so. That is the check the rest of this page cannot run on itself.