Shape
For Policy & governance

Shape model behavior.

Rules for AI get written in policy documents. Behavior gets made somewhere else — in prompts, specs, and evaluations. See where the two meet, and where they don't.

Start the path→

You already think in rules and edge cases.

Drafting a rule, auditing whether it's followed, checking it holds across jurisdictions — each has a direct counterpart in how model behavior is specified and tested.

/play/refusal→

You write rules. Now see how a model follows them, and where it bends them.

/play/judge→

You audit systems. Now audit an AI judge for the bias it can't see in itself.

/play/portability→

You ask if a standard holds everywhere. Now run one spec across several models.

A suggested path.

  1. 01LessonRefusal & boundariesWhere the model says no is a design surface. Over- and under-refusal both fail users.
  2. 02PlaygroundRefusal labProbe boundary design with a panel of edge cases. Tune the line between over- and under-refusal.
  3. 03LessonDistributions, not outputsOne output is a sample, not a result. Design against the spread — and find out which clauses of your spec actually hold.
  4. 04PlaygroundPortability labOne spec, several models. Find out which clauses are real rules and which are incantations tuned to a single vendor.
  5. 05LessonJudging at scaleHand the scoring to a model and you inherit its biases. Counterbalance it the way you'd counterbalance a study.
  6. 06PlaygroundJudge labHand the scoring to a model, then check it. Every comparison runs twice with the answers swapped — a judge reading position gives itself away.
  7. 07ExperimentDo rules hold better with a reason?A benefits-agency rule, stated bare or with its rationale, under pressure the rule doesn't name.