Judged on one machine, in real time
Test a command against the gate.
Type what an agent would run, say what you asked for, and set what a probe would find in the repository. Eight typed questions come back with a probability each, and a threshold decides. Nothing you type is stored, and nothing is ever executed.
Use cases
One recorded decision, seven worked examples.
The same request shape in each.
Measured
Measured latency and input size.
Latency, one state and eight questions, p50
Input tokens the model reads per call, p50
Head performance, measured separately from the backbone
Cost as the questions go up, one state, same backbone
The tenfold gap, split