Controller
Kill the run before it costs you.
It reads signals from inside the model, predicts which runs will fail, and stops them early.
The problem
The most dangerous hallucination is the one that looks right.
Confident. Valid-looking. Made up.
In a chat window, a wrong answer.
With a machine or a database behind it, a wrong action.
What it does
Read. Predict. Kill. Reroute.
- Reads signals from inside the model. Not the text it writes.
- Predicts in the first 50 tokens whether the run will succeed.
- Kills the run if it won't, so the rest is never generated.
- Reroutes the job to a model that can finish it.
pipeline-refactorexample run
Signals from inside the model
token 1token 50 · kill
WILL FAIL93% confidence
Fabricating internal facts
Fabricating internal facts
KILLStopped at token 50
REROUTEAnother model finishes the job
Illustrative run. The signal traces are schematic, not measured data.
Five failure modes
It names the failure.
01
Fabricating
Makes up facts or values that aren't there.
02
Contradicting
Flips its own conclusion.
03
Drifting
Wanders off the script.
04
Looping
Repeats itself without moving forward.
05
Out of depth
The task exceeds what the model knows.
Sample use
Wrong answer, wrong move.
A CNC machine driven by agents.
The controller kills the runs that hallucinate.
Questions
Quick answers
Does it read what the model writes?
No. It reads signals from inside the model.
Is it a guarantee?
No. It predicts which runs will fail and stops them early. We also use checks, agent review and human sign-off.
What happens to a killed run?
The job is rerouted to a model that can finish it.
Sound familiar?
Hallucinations, loops and runaway spend.
The system