What an EU AI Act audit actually asks for
The eight artefacts supervisors request first, and why retrofitting them after go-live costs four times more than producing them during delivery.
9 min readInsights
Field notes from programmes that are already in production.
No trend pieces. Practitioner notes from the delivery leads running governance, evaluation and adoption inside regulated enterprises.
Evaluation harnesses beat demo days
A pilot without a scored regression suite is an opinion. How to build a harness your risk function will accept as evidence.
7 min readInference cost is an architecture decision
Routing, caching, distillation and context discipline routinely take 40–70% out of run-rate without touching output quality.
6 min readWhy pilots die in quarter two
Adoption collapses when nobody owns the process the model changed. The operating-model moves that prevent it.
8 min readLineage is the cheapest insurance you can buy
When a model output is challenged, lineage decides whether you answer in an hour or in a month.
5 min readKill criteria as a portfolio instrument
Stopping two of five pilots early is not failure — it is how the remaining three get funded properly.
6 min readWant these as a quarterly briefing?
We run a closed-door session for delivery and risk leaders each quarter.