Nullius in verba · take nobody's word for it
Know when your world model is wrong, and why.
Scientists increasingly rely on AI models to predict how experiments and physical systems behave. We build tools that check those predictions against real measurements, and explain why they disagree when they do.
The problem
When a model disagrees with the measurements, there are three explanations.
In 2011 the OPERA experiment measured neutrinos travelling faster than light. The cause was a loose fibre-optic cable. Telling these cases apart is slow, expert work, and it gets harder as learned models replace hand-built ones.
01
The model is wrong
A setting changed or the model never learned this regime. Re-fit it, retrain it, or stop trusting it here.
02
The instrument is wrong
A sensor froze, drifted or picked up noise. Fix the hardware before believing any result.
03
Something new
A force or effect the model has never seen. This is the case science is looking for.
First result · Tellurion Referee
A model's own confidence is not a safety check.
On a locked test of 1,000 simulated pendulum runs, measured once, at a similar false-alarm rate:
of real problems missed by the model's own confidence
missed by the referee, which also names the likely cause
How we work
Open by default
Benchmarks, test protocols and reference code are public, so anyone can rerun our numbers.
Locked tests
Results come from a held-out test, run once and logged. We publish our weak numbers too.
Physics first
Known laws and instrument physics turn "something is off" into "this is what is off".