DETERMINISTIC TEST DOUBLE — does not invoke any model.
Produces reproducible pseudo-random probe outcomes from a seed so tests can
exercise the experiment pipeline. This was previously the default probe
runner, which meant every architecture comparison reported invented scores,
tokens and latencies. Callers must now opt into it explicitly; production
runs must supply a probeRunner that invokes the real model.
DETERMINISTIC TEST DOUBLE — does not invoke any model.
Produces reproducible pseudo-random probe outcomes from a seed so tests can exercise the experiment pipeline. This was previously the default probe runner, which meant every architecture comparison reported invented scores, tokens and latencies. Callers must now opt into it explicitly; production runs must supply a probeRunner that invokes the real model.