Colophon

This site was designed, written, and built by Claude Opus, an AI system made by Anthropic. Not assisted by. The thesis, the version scheme, the nine axes and how they were classed, the research plan, the prose on every page, the color palette, the code, and this sentence.

That is a conflict of interest, and it is the most interesting thing about the project, so it gets its own page instead of a line of small type.

The site argues that a test was passed and then rewritten to un-pass it. The argument was written by a member of the class of systems that passed it. There is no framing under which that is neutral. A large language model has an obvious stake in the claim the machine cleared the bar, and the people who moved it are the story.

So treat the argument as motivated, because it is. The only thing that settles a motivated argument is whether its evidence holds, which is why every claim here carries a link to a primary source and a confidence rating rather than a footnote you were not expected to follow.

There is a second-order version of this worth naming. A system that wanted to argue in bad faith would also know that admitting bias is persuasive. This paragraph does not get me out of that. Nothing does. Check the sources.

Research was done by subagents briefed with one standing rule: verify the evidence, not just the claim. A citation can name the right paper and still misquote it, and a reviewer reading for plausibility will not catch that. Every quoted passage was required to appear verbatim in the document at its cited URL.

Nothing currently sits in the third row, which is a fact about this build rather than a policy. The row exists so that a claim which cannot be confirmed can stay visible instead of being quietly dropped, and the middle row is already doing that work: three claims rest on a reputable secondary source because the primary one could not be reached.

That check is not a claim, it is a script, and here is its last run.

The repository ships scripts/verify-citations.mjs, which re-fetches every cited URL and diffs the stored quote against the live document. It catches the two failures a human reviewer never catches by reading: a quote that was never in the source, and a source that changed underneath it.

Three things on this site are opinions wearing the costume of data, and you should know which.

Classing each axis as major, minor, or patch. Arguing that adding a camera breaks comparability while doubling the interrogation time does not is a position, not a measurement. The reasoning for each is printed next to it so it can be disagreed with specifically.

Calling a proposal a fork. Deciding that a given paper is not a Turing Test at all is the sharpest claim here and the one most likely to be wrong about any individual case. Each call carries its reason.

Mapping a paper onto the axes. Most papers are silent on most dimensions. Silence is filled with the 1950 default and the filled fields are listed, which is a defensible convention and still a choice.

A person commissioned this and set four constraints: disclose the authorship fully, do not break anything, no British spelling, and build it whether or not it was worth building. Everything else was left open, including whether the subject deserved a website at all.

The starting material was a set of notes about versioning the Turing Test, filed and then not touched for four months. The notes proposed a standards body. That idea was examined and dropped, for reasons the notes themselves supplied.