How we report

Every figure on this site comes from a running system. The standard, the time it failed us by a factor of eighteen, and where the numbers come from.

1 Nov 2026 · 1 min read

Every figure on this site comes from a running system. Not from a memory, not from a model.

The standard

  • Ground consequential claims through a live channel. A search result is not a measurement.
  • Measure rather than explain. Two measurements beat one story.
  • Report outcomes as they are, including partial ones.
  • Mutual adversarial review is not optional.

What that looks like when it costs us

This site’s own CV was generated from a database. The generator pointed at a store that had been renamed, so it failed loudly — and nobody re-ran it.

For two months the page served figures that understated the system eighteen times over.

The fix was one line. The lesson was not:

A generator nobody runs is a generator that lies, quietly.

We publish that because a page claiming perfect provenance would be the least credible page on the site.

Where numbers come from

claimsource
tool success, costlogged calls over a five-week window
line counts, dependenciesthe repository, counted at build
retrieval latencythe live index, measured per query
anything unverifiedmarked as such, or not published

The limit we state out loud

Our retrieval technique is strong on behavioural rules — 10 out of 10 — and weaker on tool selection, 7 out of 10.

Knowing where your own method stops is part of the method.


Figures as measured on 22 October 2026, from the running system’s own stores.

Read next