Blog
Field notes from the action layer.
Injection denials, honest abstention, and signed evidence — written up as we build.
2026-07-21 · 1 min read
The agent lied. The gate didn't care.
Why enforcement must key off confirmed model state, not agent self-reports, in the era of 'assume breach'.
Read the post →2026-07-20 · 1 min read
Zero wrong answers with the model unplugged
What happened when we ran our entire NLQ test suite with no working LLM at all.
Read the post →2026-06-01 · 1 min read
Why AI Agents Need a Semantic Trust Layer
Column names are not business definitions. Here is why every AI agent over enterprise data needs Epistom.
Read the post →