FIELD NOTES FROM THE BOUNDARY

Insights

Engineering notes from systems where the enemy is a wrong number that looks right. Everything here comes from shipped work; nothing is hypothetical.

001 10 AUG 2026 · 6 MIN · LLM PIPELINES
Ratio checks can’t catch a two-row drop

A model returned 52 of 54 rows and every plausibility check passed. Why statistical checks structurally cannot catch silent data loss — and why the fix was a set difference, not a better prompt.

002 10 AUG 2026 · 7 MIN · EU AI ACT
The EU AI Act wants your logs to prove something. Can they?

The high-risk obligations apply since 2 August 2026. An engineer’s reading: record-keeping, traceability, human oversight and accuracy are architecture properties — and five questions that tell you where you stand.

003 10 AUG 2026 · 5 MIN · LLM INTEGRATION
Schema strictness should track consequence

The same system validates some model outputs with forced schemas and others with a shrug — and that’s correct. A taxonomy for matching validation contracts to blast radius.

004 10 AUG 2026 · 5 MIN · ARCHITECTURE
Refuse to compute

The most underrated feature in a regulated system is a clean no. A pre-flight rejection is a support ticket; an approximate answer is an assessment — for tax engines and AI agents alike.