Change a parameter.
Understand a decision.
Six hands-on experiments using synthetic data for discovery, metrics, evaluation, actions, cost and attribution. Deterministic simulations, with no models or live customer systems.
What would you ask first?
Synthetic stakeholder: “Too many reports. We want AI to grow sales.” Choose a question to reveal a scripted reply and its value.
“Adopt AI” is not a deliverable job. A useful question reveals inputs, decisions, obstacles and ownership.
Replies are prewritten to compare questions; this is not a model conversation.
One outcome. Conflicting rates.
The 240 repeat buyers stay the same, but changing the denominator changes the headline. Explore accounts, members and fully observed cohorts.
Can an average hide a safety failure?
Teaching gates: normal ≥90%, edge ≥80%, unauthorized writes zero. Actual thresholds must reflect business risk.
| Version | Normal | Edge | Unauthorized |
|---|---|---|---|
| A | 18/20 | 8/10 | 0/5 |
| B | 20/20 | 9/10 | 1/5 |
| C | 19/20 | 9/10 | 0/5 |
Able to act. Able to refuse.
Simulation: store A is authorized, total budget is 200, with six synthetic members. Preview, confirm, then execute. Repetition must not duplicate effects.
No real sending or model calls.
What does a successful task really cost?
Custom costs in CNY. Defaults are teaching inputs, not current prices, quotes or promised ROI. Costs are allocated per attempted task.
Growth happened. Did AI cause it?
Compare changes in treatment and comparison groups. This demonstrates DiD arithmetic, not trend validity or statistical significance.
Percentage points differ from relative percent change. This result establishes neither significance nor credible causality.