Sets how much an AI system remembers and how much it varies, then runs a Bias Audit that produces evidence, not opinion.
Private skill · shared by invitation · How access works
More context is assumed to be better, and it isn't always.
Memory and randomness are independent, and configurations mix them up.
Vendors imply knobs the platform doesn't have.
Policies describe fairness without testing it.
Map the dial to the controls your platform actually exposes.
Four starting presets across memory and variance, with the fabrication risk stated.
The same decision at full, partial and no context, matched pairs, five or more runs.
What moves when context is stripped is what the context was doing.
The AI system, its configuration, and the queries that matter to the business.
Deciding what level of variance the business will accept.
Thirty minutes with Harveer, on your own data problem.