Journal
Dated records rather than claims: what changed, what real use taught us, and what the web let through.
Everything in this section carries a date, and every date can be checked against the repository history. That is the point of the section: a claim you cannot contradict is not worth publishing.
These are internal working records, served as they were written. They use the project’s own vocabulary — phase and file names that mean nothing outside it — and they have not been rewritten to read better from here. A log tidied up after the fact stops being evidence.
On 26 September 2026: 65 dated entries in the development log, 71 frictions in the field notes at their last cumulative count, 53 published versions, and two measured access campaigns.
What each record is for
The development log records decisions, not commits — why something changed that way, and what the alternative cost. It is written during the work, which is what makes it occasionally awkward: the reversals are in it too.
The field notes are an operator’s log, in French, kept in the language the work was done in. They are the raw material the recipes are distilled from.
How models behave records what language models actually do with this tool, including what they do wrong — an entry from 3 July 2026 describes a model driving Diwall for several sessions without reading its guide, and the lock that now prevents it.
Web accessibility counts what public pages answer to an agent. On 27 June 2026: 39 % blocked, 8.7 % accessible. The full breakdown, with what it is missing →
Three failures, and what caught them
The journal keeps the misses along with the wins: they are what shows where the tool breaks.
- A credential shown in clear text during a debug session, spotted by re-reading the session output — which is what led to the Trust page being rewritten.
- Stealth mode (
--stealth) reported as active in the boussole while it was not applying, ever since it shipped: a silent API break had degraded it to a no-op. Found while trying to measure its effect on a benchmark. - A translation run that returned a paragraph with its meaning inverted (8 September 2026), caught by back-translation and the semantic diff.