Laboratory
Every figure, with the command that produced it.
This site preaches measurement discipline. So it applies it to itself: these figures are regenerated by a script before every publication, never typed by hand. The command is shown next to the value.
measured on 2026-08-03
| Value | What it is | How it is measured |
|---|---|---|
| 1,162 | commits | git rev-list --count HEAD sur 17 dépôts |
| 17 | repositories | inventaire des dépôts de l'écosystème |
| 5,472 | declared tests | comptage des `def test_*` dans les suites pytest |
| 1,699,549 | indexed chunks | SELECT count(*) FROM chunks — Library Brain |
| 19.5 | GB RAG database | taille du fichier SQLite |
| 122 → 0 | open CVE declared | pip-audit sur 4 environnements de production |
| 0 | cloud dependency declared | par construction — aucun appel sortant |
| 0 | third-party request on this site declared | vérifiable : DevTools › Réseau, sur cette page |
| 208 | GB of redundant RAM removed declared | poids cumulés avant/après mutualisation par le gateway |
| 8 | public MIT repositories | gh api user/repos --jq 'select(.private == false)' |
Why this page exists
On 3 August 2026 this site claimed 1,039 commits while measurement gave 1,150. The problem was not the gap — it was that the number was typed by hand, and therefore doomed to go stale. A figure you cannot regenerate is a figure you cannot trust.
Since then, tools/collect_metrics.py measures before every site build, and tools/check.py refuses to publish if the measurements are older than the content.
The lines marked "declared"
Some values are not produced by a command on every build: the number of CVEs cleared is a historical result, the gigabytes of RAM saved come from a before/after comparison. They are marked as such rather than presented as automatic. A declared figure that says so beats a declared figure in disguise.
What is not shown
Lines of code, deliberately. It is a vanity metric: an experienced engineer is judged on the lines they remove and the abstractions they avoid, not on volume produced. The number of integrated models is missing too — it is plausible, but I do not yet have the command that produces it, and a figure without a command does not go on this page.
Your data cannot leave the building?
That is precisely the problem I solve. A 30-minute call is enough to scope an audit.