How the teaching model, the AI's guardrails and the 16 domains work together. For the short version, see the main Northstar page.
| Part | What it does |
|---|---|
| Learner inputs | Share of each group in the training data, decision bar, factors on or off, algorithm |
| Scenario data | Factors, how strongly each one stands in for a group, group labels, real-world grounding |
| Teaching model | Produces every number on screen. The same setup always gives the same numbers. Labelled "teaching numbers, not real measurements" |
| Outputs | Approval rate per group, fairness gap, what drove it, next steps |
| AI narrator and study material | Receives the exact numbers. Explains, quizzes and answers follow-ups |
AI guardrails: never invents numbers, plain language, hedged claims, no verdicts.
The first concept had four scenarios. I grew it to sixteen so learners could find a case close to their own life. Learners can also upload their own CSV, which stays in the browser.
Every domain uses the same template, so the model stays identical and only the story changes.
| Promise | See how an AI screens job applicants for interviews |
|---|---|
| Outcome | "approved for an interview", used in every label and sentence |
| Groups | Men and women on the gender lens. Race is the second lens |
| Teaches | A factor that looks neutral, like past pay, can carry a group's history with it |
| Real world | Real résumé screeners have down-ranked applicants linked to women's colleges and clubs |
| Factors | Experience (no link to gender), education (weak stand-in), age (no link to gender), previous salary (strong stand-in) |
| Machine-learning term | What learners see |
|---|---|
| Dataset composition | How balanced the training data is |
| Decision threshold | How strict the decision bar is |
| Feature selection | What the AI is allowed to look at |
| Proxy variable | A factor that stands in for a group |
In health, justice, benefits, housing and livelihood domains, the AI narrator: