25 June 2026
Why 'temperature zero' still isn't determinism.
Sampling settings are not an architecture. What it actually takes to re-mark a script months later and get the same grade, byte for byte.
Full post coming to this space →
News · AI news
AI news.
What's changing in machine learning that actually affects regulated assessment — separated from what's just noise.
3 June 2026
Dozens of scripts, not thousands.
The industry assumption is that useful marking models need enormous labelled corpora. Schema-driven extraction changes the arithmetic.
Full post coming to this space →
8 May 2026
LLM judges are easy. Auditable judges are not.
A judging model is a weekend project; a judging system an EQA can interrogate is the actual work. Where the gap lives.
Full post coming to this space →