Where we are
Nine sessions of pieces: the workspace and house rules (b1), scorecards and job ads (b2), sourcing (b3), the screen you fixed from a 0.09 four-fifths ratio to 1.00 (b4), structured interviews and calibration (b5), offers and decision records (b6), onboarding (b7), the performance cycle (b8), and the skills map and attrition read (b9). Tonight they stop being nine exercises and become one quarter. The point of the capstone is not more technique - it is the joins, because that is where real HR functions drop things: the impact check nobody ran because hiring was busy, the decision record written a month late, the rejection sent by a template that cited a criterion nobody applied.
Everything, in one picture, with two checkpoints 6 min live
Ten stages from scorecard to onboarding, and two places where the process stops until a human does something. Everything else can be drafted, templated, connected and sped up. These two cannot, and a quarter that skips either one is not fast - it is undocumented.
LiveThe two checkpoints, and why these two3 min▶
Everything else in the pipeline is craft. These two are the difference between a process you can explain and one you can only apologise for.
- The impact check before any slate is acted on. Selection rates per monitored group, the four-fifths ratio, and a stop if it is under 0.80. Run it per slate and again across the wave, because twelve seats swing a small ratio hard while the aggregate tells you whether it is a pattern. Remember what b4 taught: the check itself fixes nothing, it just makes a silent failure visible in time to fix it.
- A named decision-maker with a reason, written the same day. Every shortlist, every rejection, every offer, every rating. Not because a form demands it, but because the challenge arrives months later and a rationale reconstructed after the objection reads exactly like one.
The quarter that was fine until anyone asked. A team ran a clean process on paper: good scorecards, structured interviews, calibrated raters. Then a rejected candidate asked why, eleven weeks later. The criteria existed. The interview notes existed. What did not exist was any record of who made the call or what tipped it, because the decision happened in a corridor between two meetings and everyone remembered it slightly differently. Nothing unlawful occurred. There was simply nothing to show, which in practice is the same problem.
Self-studyWhere the joins usually break2 min read▶
- Scorecard to screen. The criteria that get screened on quietly stop matching the ones that were agreed, usually because the screen needs something machine-readable and years of experience is easy.
- Screen to interview. The panel has not read the scorecard, so the structured questions get abandoned in the room for a nice chat.
- Interview to decision. Evidence exists per rating, and then the decision gets made on the general feeling of the debrief.
- Decision to record. The record is written when someone remembers, which is when someone complains.
- Hire to onboarding. The scorecard that defined the role never reaches the manager, so the 30/60/90 measures something else entirely.
Every one of these is a handover between two people who are both busy. That is what the capstone run is for - not to practise the stages, but to practise the seams.
Run Kestrel's quarter ★ 16 min · everyone builds
Short prompts, because the workspace already knows Kestrel. Three artifacts get built live here; the other two are Demo 1. Notice how little instruction each one needs now compared with session b1 - that is what the context pack was for.
Monday - refresh the pack. Update the roles filled, the open slates, the risk of the week, and the last cycle's outputs. Five minutes, and everything downstream depends on it.
Tuesday - the wave impact check. Aggregate selection rates across every role filled this quarter, the four-fifths ratio, and the per-slate view underneath it. If it fails, the escalation from leader session a3 fires today, not next month.
Wednesday - the hiring-manager update. One page per manager: where their roles are, what is blocked, what you need from them, and by when.
Thursday - decisions and records. Offers, rejections, and every decision record written the same day by the person who decided.
Friday - the cycle pass. The performance-cycle evidence pass from b8, and the flagged list of claims with no artifact behind them goes back to managers.
What you never delegate, and why 5 min live
Not because a model cannot produce words for these, but because in each one the accountability is the substance. Something is at stake for the person on the other side, and nothing is at stake for the machine.
LiveWhy each one, in a sentence2 min▶
- The decision about a person. It is regulated, it is consequential, and being accountable for it is the entire job. A ranked list is an input; the decision is a person's act.
- The bad news. A rejection or a hard rating delivered by something with nothing at stake reads exactly like what it is. The cost of sending it is part of what makes it honest.
- The conversation where someone is struggling. Illness, burnout, a grievance, a resignation. There is no draft for the moment somebody tells you the truth.
- The judgment about the bar. What "good enough" means at your company is a values decision that shapes who gets in and who gets ahead. A model will happily infer it from your history, which is precisely the thing you are trying not to repeat.
The audit file 3 min live
Not a compliance exercise for its own sake. If you can produce these rows on request, your process is real; if you cannot, the quality of your intentions is not going to help. Any missing row is this quarter's work, in priority order.
| Row | What it is | Built in |
|---|---|---|
| Criteria | The scorecard per role, with each must-have evidenceable | b2 |
| Slates | Who was in the pool at each stage, not just who came out | b3 · b4 |
| Evidence | What supported each rating, per candidate, per criterion | b4 · b5 |
| Decision records | Who decided, their reason, the date it was written | b6 |
| Impact checks | Selection rates and the four-fifths ratio, per slate and per wave | b4 · tonight |
| Notices | What candidates were told about AI use, and when it was sent | b1 |
| Tool versions | Which tool and version was in use for each decision period | b1 · leader a3 |
Six rows out of seven. A people team ran this inventory expecting to fail it and found they had six of the seven already, scattered across an ATS, a shared drive and somebody's inbox. The missing row was tool versions, which nobody had thought to record - and which mattered, because their vendor had shipped a model change mid-quarter that nobody could date afterwards. Producing the file took a fortnight. Not producing it would have taken an incident.
The two remaining artifacts ★ 10 min · everyone builds
You built three in Part 2. These are the last two - one complete decision record, and the audit file index for the quarter.
Take one filled role and assemble the complete decision record: criteria, slate, evidence per rating, the named decider and their reason, the impact check result, and the tool in use.
Read it as if you were the rejected runner-up. Anything you would find unanswerable is a gap.
Build the audit file index for the quarter - the seven rows, where each lives, who owns it, and which are missing.
Rank the missing rows by how long they would take to start producing. Start with the cheapest, this week.
Your own 30 days ★ 8 min · build your own
Kestrel is fictional. Your Monday is not. Four weeks, one artifact each, in the order that gets you defensible fastest rather than the order that is most fun.
Week 1 - the workspace and house rules. Six documents, none of them personal data, plus your never-list. Two hours, and everything else gets easier.
Week 2 - one role scorecard, done properly. Every must-have evidenceable. Use the role you are hiring for right now, not a hypothetical one.
Week 3 - the impact check on your last five roles. Aggregate, monitoring data only. Whatever the number, you will know something on Friday that you did not know on Monday.
Week 4 - the screening protocol. The six levers from b4 written down, with who runs the check and what happens when it fails.
Pick your measures now, before you start: hours moved off drafting, share of criteria that are evidenceable, share of decisions with a same-day recorded reason, and the four-fifths ratio across the wave. Time-to-hire on its own is the one to distrust - it falls fastest when you skip the checks.
Three rails, and a habit 2 min live
Ten sessions come down to three sentences and one discipline.
- A person decides, and the record shows which person - with the reason written the same day, not reconstructed after somebody objects.
- Rate the job, not the polish - an undefined criterion does not stay neutral, it quietly measures background.
- People data is not prompt fodder - candidates and employees did not hand you their information so it could become context.
And the habit: check the current position before you act. The EU AI Act's employment obligations moved from August 2026 to 2 December 2027 while this course was being written, and NYC's rules were live years before most teams noticed. Anything in these pages could move again. The craft - structured criteria, evidence, calibration, a human decision, a real record - does not.
Try it yourself - the next 30 days ◐ one artifact per week
- Run week 1 this week. The workspace and house rules are the prerequisite for everything else in the track.
- Run the impact check on your last five roles in aggregate, and write down the number before you interpret it.
- Build your audit file index and count the missing rows. Start producing the cheapest one immediately.
- Write your own never-delegate list and share it with your team, so the boundary is not just in your head.
- Diary a quarterly re-check of the legal position with whoever owns compliance. The habit is the deliverable.
Official sources covered
The capstone draws on every source in the course. This is education for HR professionals, not legal advice, and obligations vary by jurisdiction - verify the current position before you act. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Which two points in the pipeline stop the line until a human acts?
Everything else can be drafted, templated and connected. These two are what make the quarter explainable afterwards - and a process that skips either one is not faster, it is unevidenced.
2 · Your wave-level impact check passes at 0.86 but one individual slate came in at 0.50. What does that tell you?
One seat moves a twelve-person ratio a long way, which is why you run both. A failing slate inside a passing wave is a prompt to examine that slate's criteria and raters - a signal that starts an investigation, not a verdict on its own.
3 · What is the test for whether something belongs on the never-delegate list?
The decision, the bad news, the hard conversation and the bar - in each one the accountability is the substance, and the model has nothing at stake. The legal question is narrower than the right question, and drafting quality is not the point at all.