Invariant-based verification for enterprise software agents
A design framework for verifying task completion against business constraints and testing verifiers adversarially.
SimReal Research
Research
A logbook of discoveries from building worlds that teach intelligence.
The thesis, in four lines
01
Chess, Go and code had a world that answered back.
A winner. A failing test. A way to try again. That is what let them fall to AI.
02
Beyond games and code, the answers are hard to read.
They arrive months late, mixed with luck, and only once.
03
A judge built by the model gets gamed.
An optimiser finds the cheapest path to the number. Only the real rules stay out of reach.
04
So we build worlds that answer back.
The same situation, replayed. One decision changed. The real rules determine what follows.
Guides
Start here
Plain-language explainers on the ideas behind our work.
Publications
In progress
SimReal Research