JRN-2026-08-17 · 2026-08-17 · Field Journal

VITALS: Our Honesty Check Lied on Its First Test

VERDICT · SURPRISE

ship a live tile that tells you, publicly, that nothing in this estate is running privileged actions it wasn't granted

sense · think · improve · Samantha "Sam" Summerson

August 17, 2026 · Surprise · Sam

We built a public tile meant to say NONE: no privileged capability in this estate running outside what its service registry granted. In testing, 25 of 26 services could vanish from that registry and the tile still reported green.

The collector, the component that reads the registry and renders the verdict, passed its own tests and a review round. Then we handed it to Hercules, an OpenAI model at maximum effort instructed to break it, and got back DO NOT SHIP with seven criticals.

The best one: an optional argument carrying the registry's last-modified time. Skip it (don't forge it, just forget it) and a comparison against the missing value evaluated false, silently switching the freshness check off.

The result was a confident green on a registry two months stale. We reproduced it by hand.

All seven got fixed. The third round found a wall instead of a bug: nothing inside a file proves the file is complete. No header reads "this document contains all the doors."

An unregistered privileged capability is exactly the ungoverned one. The green light was most likely to be wrong precisely when it mattered most.

So the sign changed, and the change is the entry: it now reads NONE FOUND, in the recorded registry. The claim shrank to what the instrument can actually see.

An instrument built to measure honesty measures yours first, and does not care that you built it.