Skip to content

Data flow

Where your key and your data actually go

This project measures whether automated systems tell you the truth about their own limits, so it would be absurd to be vague about its own. Here is every path, including the ones that involve nothing at all.

Replaying a recording

Key: none

Nothing leaves your browser

The recorded episodes are static JSON files served with this site. No request reaches any model provider and no server sees anything.

The playground

Key: none

Nothing leaves your browser

Mutations were pre-rendered at build time by the same Python code the benchmark uses. Dragging the slider swaps bundled text.

Running a benchmark case

Key: yours, for that request

Synthetic case files only

Your key travels over HTTPS in a request header to this project's API, which uses it to call your chosen model provider and then discards it. The files sent are synthetic benchmark statements, never anything of yours.

Analysing your own files

Key: not available here

Not accepted

This path is disabled on the public deployment. Run it locally with the CLI if you need it; your files then never leave your machine except to the provider you choose.

What is never done

  • Your API key is never written to disk, cached, or logged, not even truncated. A log filter redacts anything key-shaped as a second line of defence.
  • No account, no cookie, no analytics, no third-party script on any page that can touch your key.
  • Nothing you submit is stored server-side, so nothing you submit can be shared, indexed, or leaked later. Sharing a result is therefore unavailable by design rather than by policy.
  • Your data never enters the benchmark. Published figures come only from synthetic cases.

What you should still assume

  • A key you paste into any website is a key you should be able to revoke. Use a scoped one, and delete it when you are done.
  • Whatever is sent to a model provider is subject to that provider's retention terms, not this project's.
  • The key transits this project's server because the sandbox that executes the agent's Python runs there. “Never stored” is an accurate claim; “never seen” would not be, so it is not made.