Institute for Agentic Research · live
Independent research on agentic AI.
In plain words
We built one AI agent, used it as our single test subject (n=1), and carefully switched off pieces of it one at a time to learn what each piece actually does. Before each test we wrote down what we expected — and we published every result, including the prediction that turned out wrong.
n=1
n=1 means one test subject. We make no claim that what we found generalizes — we report what happened to this specific system.
Ablation
An ablation removes one component to see how the whole behaves without it — the same way medicine learns what an organ does.
Pre-registered
Pre-registered means we wrote our predictions down and hashed them before running the experiment — so we can't quietly rewrite history if a guess turns out wrong.
In numbers · 2026
One production system, five subsystems , ablated one at a time — with each prediction registered before the data came in.
0/90
Score · architect rating · upper
0/0
Ablations hit pre-registered targets
0%
H2 observed · pre-registered ≥60% — failed, reported
n=1
Honest sample size — no generalization claim
Earlier writing
FRANK · 30 Mar 2026
Your AI Doesn't Know You. SOMA Changes That.
Every chatbot talks to you the same way it talks to everyone else. We built a system that silently rewires a 3-billion-parameter brain while it sleeps — so yours doesn't have to.
Gabriel Schaider
RESEARCH · 30 Mar 2026
A 3-Billion-Parameter Model Just Diagnosed Its Own Bug. We Checked. It Was Right.
At 20:54 on a Tuesday evening, Frank — a local AI running on a laptop in Austria — produced an idle thought that correctly identified a processing flaw in his own cognitive architecture. Nobody asked him to. He just... noticed.
Gabriel Schaider