Why Models Behave Differently: Parameters, Variability, and Control
Participants hold a task steady, vary one condition and repeat trials. They use the evidence to decide when variation helps, when it creates risk and which checks are needed.
Literacy perspective
Mechanism Literacy
How tokens, attention, parameters, sampling, variability, and model behaviour affect predictability.
What lecturers should notice
Use this as an experimental lens
Separate observable behaviour from speculation about hidden internals, then manipulate generation conditions and repeat runs.
Related units
2 units make this literacy visible
The role may be primary, secondary, supporting, integrated, or part of the attack surface; consult each unit’s metadata for the exact relationship.
Reflection Loops: Critique, Revision, and Better Reasoning
Participants organise critique, evidence and human decisions into workflows. In the Arena they first improve a judge, then keep the selected judge fixed while testing a generator change.