The dossier
The long read, for the reader who checks.
Ten sections, in reading order, each with what stands behind it. Lumen is written to be checked, and this page is the map.
- Edition
- 2026.08
- Sections
- Ten
- Languages
- Three
- Written
- Ottawa
Foreword
Two kinds of numbers appear on this site. The first kind we measured ourselves, and each one is stated with its method and its hardware. The second kind was published by others, and each one is stated with its study, its year and its cohort. The two are never mixed, and neither is ever rounded up.
Read the sections in order and the system assembles itself: the people, the clones, the iron they run on, the jurisdiction they answer to, and the signatures that govern them. Start anywhere and each section still carries its own sources. The status of everything, stated plainly, lives on the investor page.
Ottawa · August 2026
Table of contents
The two kinds of numbers
Measured by us
Compression on the lossy pass of a real 31-billion-parameter model, 58.25 GB to 5.92 GB.
Internal measurement · reference hardware
Tensors recovered bit-identical in the separate lossless pass. The two results are never combined.
Internal measurement · reference hardware
Published by others
of claims in a national US malpractice database involved a communication failure. A finding about hospitals in general, not about Lumen.
Humphrey 2022 · published study
fewer preventable adverse events after structured handoffs in a trial across nine pediatric residency programmes.
Starmer 2014 · published study
The standing rule behind both columns: no clinical outcome number of our own appears anywhere on this site until one is measured at a reference site.