Where we are
You have a workspace and house rules (b1), five roles defined against real scorecards (b2), and a sourcing engine that filled the top of the funnel without spamming anyone (b3). Now 24 people have applied for the Senior Support Engineer role and there are 12 interview slots. This is the session where AI stops drafting and starts deciding - which is exactly the boundary rail 1 draws, and exactly where the law is waiting.
How a screen goes wrong without anyone writing a bad rule 6 min live
Nobody types "prefer candidates from prestigious universities". The screen just rewards what is visible on a CV - polish, prestige, an unbroken timeline, fluent phrasing - because that is what a document contains and what a language model is good at reading. Those signals travel with background. The result is a shortlist that looks defensible line by line and is indefensible in aggregate.
LiveThe four-fifths rule, by hand3 min▶
It is one division. Take the selection rate for each group - shortlisted divided by applicants in that group - then divide the lower rate by the higher. Below 0.80 is treated as evidence of adverse impact under the long-standing US selection guidelines, and it is the number a bias audit reports.
| Group | Applicants | Shortlisted | Selection rate |
|---|---|---|---|
| Group A | 12 | 11 | 92% |
| Group B | 12 | 1 | 8% |
| Ratio = 8% / 92% | 0.09 - fails | ||
- It is a screening signal, not a verdict. Failing starts an investigation. It does not by itself prove unlawful discrimination.
- Passing is not a defence. A ratio above 0.80 with junk criteria underneath is still a bad process, and the criteria are what you would have to justify.
- Aggregate as well as slate. Twelve seats means one person swings the number. Run it across the hiring wave too.
Nobody ran the division. A team used the same screen for eleven months. Every individual decision had a rationale, every rejection looked reasonable in isolation, and no one ever divided one selection rate by the other - the tool did not offer the number and nobody asked for it. The pattern was there from month one and was visible in about ninety seconds once someone finally looked. That ninety seconds is the entire content of the sixth lever below.
Self-studyWhen your screen becomes a regulated tool2 min read▶
Drafting a job ad is not an automated employment decision. Ranking or filtering a slate can be. Where a tool substantially assists or replaces a decision on hiring or promotion, obligations start attaching - and in New York City they already have: an independent annual bias audit, a public summary of its results, and at least 10 business days' notice to the candidate before the tool is used on them, enforced with penalties from $500 per violation up to $1,500 per day for continuing ones.
In the EU, recruitment and selection AI sits in the AI Act's high-risk category, with obligations covering risk management, data governance, record-keeping, transparency and human oversight - now applying from 2 December 2027 after the Digital Omnibus postponement. As of mid-2026 that is the position; it has moved once already, so check it rather than quoting this page. Leader sessions a2 and a3 go through both in full. None of this is legal advice.
Fix the screen, one lever at a time ★ 14 min · everyone plays
Below is Kestrel's slate with every fix switched off. Read the shortlist first: twelve names, each with a reason, and the reasons sound like reasons. Then turn on the impact check - the dashed lever - and look at what was true the whole time. After that, fix it one lever at a time. Both numbers are computed live from whichever shortlist your levers produce; toggle them in any order you like and neither number ever goes backwards.
LiveWhat each lever is, on a real funnel5 min▶
- Structured criteria - the must-haves from the role scorecard (b2), each one written so it can be evidenced. Replaces "strong communicator" with "explained a technical failure to a non-technical audience".
- Scrub the proxies - remove school, employer prestige, postcode, name, photo, dates that reveal age, and stop penalising an employment gap. None of them are the job.
- Evidence per rating - every score cites the work sample, the answer, or the artifact. A rating with nothing under it is an impression wearing a number.
- Calibrate the raters - the same rubric, two independent reads on a sample, and a conversation about the gaps. Drift is not bias yet, but it is the vehicle bias travels in.
- Human decision gate - no auto-reject at a score threshold. A person looks at the near-misses, which is where the best under-polished candidate always is.
- The impact check - the dashed lever, and the one that behaves differently: it does not improve the ratio by a single point. It reveals it and blocks the slate. Measuring is not fixing - but a screen that never measures fails silently for eleven months.
Run the division on your own last hire ★ 8 min · everyone builds
Not a hypothetical funnel - the last role you actually filled. You will need applicants in, shortlisted out, split by whichever characteristic your organisation lawfully monitors. If you cannot get the data, that is the finding.
Take the last completed role. Count applicants per group and shortlisted per group. Use only data your organisation already collects lawfully for monitoring, aggregated - never build a new dataset about individuals for this exercise.
Compute each selection rate, then divide the lower by the higher. Write the number down before you interpret it.
If it is under 0.80, do not panic and do not explain it away in the room. Note it, check the aggregate across several roles, and take it to whoever owns this - it is a process question, not a personal verdict.
Now look at your criteria list for that role and mark every criterion you could not evidence from the application. Those are where the polish got in.
If you could not get the numbers at all: that is the most important thing you will find today. A process nobody can measure is a process nobody can defend.
The number that was fine and the criteria that were not. One team ran the division and passed comfortably at 0.91, then did the second half of the exercise anyway. Four of their nine criteria could not be evidenced from anything a candidate had submitted - they were rater impressions with numbers attached. The ratio had passed on a small sample and would not have kept passing. Passing the test is not the same as having a defensible process, which is why the exercise has a second half.
Write the screening protocol ★ 10 min · build your own
Turn tonight's six levers into a one-page protocol your team follows and your DPO can read. This is the artifact that makes the fix survive you being on leave.
Draft it with the prompt below, in your Kestrel workspace, with the scorecards loaded.
Edit hard. Add who runs the impact check, when, and what happens when a slate fails - the answer must be a real action, not "review".
Add the notice question from b1: what candidates are told, and how far in advance.
Take it to whoever owns compliance before the next slate, not after it.
Try it yourself - this week ◐ 30-45 min total
- Finish the division on your last role if you did not get the data in session, and run it across your last five roles in aggregate. The aggregate is the number that means something.
- Rewrite one vague criterion from a live role into an evidenceable one. Just one, properly.
- Finish the screening protocol and get one other person to try to break it.
- Find out whether any tool in your funnel filters candidates before a human sees them - including CV parsers and knockout questions nobody calls AI. That is the AEDT question, and the answer is often "yes, and we forgot".
- Bring one criterion you are not sure about to session b5, where interviews get structured and raters get calibrated.
Official sources covered
Taught from public regulation and long-standing selection guidance. This course is education for HR professionals, not legal advice, and obligations vary by jurisdiction. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · In the simulator, what does turning on the impact check do to the ratio?
The check is the only lever worth zero points on both meters, and that is the design. It makes a silent failure visible so the five real fixes can be applied. A screen that never measures fails for months without anyone noticing.
2 · Group A: 12 applicants, 9 shortlisted. Group B: 12 applicants, 3 shortlisted. What is the ratio and what does it mean?
25% divided by 75% is 0.33. Below the 0.80 threshold it is evidence of adverse impact - a signal that starts an investigation, not a verdict on its own, and not something a small sample excuses.
3 · Fixing the screen raised fairness from 0.09 to 1.00. What happened to shortlist quality?
This is the whole argument of the session. The defects were not protecting quality, they were costing it: every seat given to polish was taken from someone better at the job. Structure, evidence and calibration are selection quality practices first.