Today's task: interview questions for one role, from a job description and a CV. Personal data in, questions asked of real people out, hard to undo. The governance question: who decided, on what evidence, and who answers for it.
Pick a token, then click a stage
HumanAINobodyor click a stage to cycle, or drag.
Nothing gets produced: put Human or AI in Generate.
Load a layout from today
What that layout does. Green is low risk, red is high.
Speed
Errors caught
Human effort
How this is scored
A teaching model, not a measurement. The live run in Exercise 07 is the measurement. The rules are fixed and simple so you can argue with them.
Speed starts at 100. A human costs 10 at the matrix, 45 at generating, 20 at evaluating and 10 at review. An AI costs 5 per stage. Nobody costs nothing.
Errors caught starts at 55 for an AI-written question or 70 for a human-written one, then adds 15 for a human-read matrix, 5 for an AI matrix, 25 for a human evaluation, 12 for an AI evaluation (format and open-endedness, not fairness in effect) and 5 for human review before the bank. If AI generated and AI evaluated and the only human is the review with no one at the matrix, that final 5 is lost: the person skims because two agents already passed it.
Human effort counts 1 for the matrix, 4 for writing, 2 for evaluating, 1 for review, out of 8. Six or more is flagged as fatigue.
Accountability is Low if a human reviews before the bank, Medium if a human evaluated or wrote but nobody reviewed, otherwise High.
Data and legal starts high, because the CV is personal data. It drops a level if a person reads the question before a candidate does and another if a person read the matrix, and is High whenever AI wrote it and no person reads it.
Security is Low with no AI stage. Otherwise it starts at how exposed the CV route is to a planted line, and drops a level for a human read before the bank and for a human at the matrix.
Quality follows the errors-caught score: under 50 High, 50 to 75 Medium, above 75 Low, one level worse because a question asked cannot be unasked.
People is Medium when most of the work sits with people (fatigue), when a human both checks and signs off, or on the rubber-stamp flag, otherwise Low. Nothing here is High: people problems show up as slowness and skipped checks, which the other rows then catch.
Acceptable or too risky: too risky if any row is High, or if more than one row is Medium. Hiring earns the stricter line.