Course › Module 5 · Candidate matching and ranking

Lab 4: rank and justify a shortlist of ten

Module 5, Lesson 5  ·  4 min read ·  Updated 21 September 2026

Module 5 · Lesson 5

Rank ten people yourself, write down why, then compare with the software. The point is whether your reasons survive being written down.

What to do

  1. Take the shortlist of 20 from Module 4.
  2. Build a match profile from your scorecard using the method in Lesson 2.
  3. Run it and note the software's order for those 20.
  4. Before looking at that order, pick your own top ten and rank them.
  5. For each of your ten, write one sentence saying why, quoting a specific line from the CV.
  6. Compare the two orders.

Step 4 has to happen before you see the software's result. This is not a theoretical risk — if you read the ranking first you will produce a lightly edited version of it and learn nothing.

What to hand in

A table with ten rows: your position, the software's position, the candidate, and your one-sentence reason with its quote. Plus two or three sentences on where the two orders differed most and why.

How to know you have done it right

CheckGood looks like
Every reason quotes somethingA line from the CV, not an impression
Reasons match your rulesEach one names something from your scorecard
You can explain the differencesYou can say why the software ranked someone differently
You used bandsYou treat 1–3 and 4–6 as groups, not a strict order
At least one disagreementTotal agreement usually means you looked first

What the differences usually mean

PatternLikely reason
You ranked someone much higherYou are using knowledge the CV does not contain. Check whether it is real or assumed
The software ranked someone much higherOften length or keyword density rather than substance
You disagree about nearly everyoneThe match profile is not built from your actual rules
The orders are almost identicalYou saw the ranking first, or your rules are too blunt to separate people

Check yourself before Module 6

  • Could you defend your number one to a hiring manager using only what is in the CV?
  • Could you explain to your number ten, in plain words, why they came tenth?
  • Did any of your reasons turn out to be an impression you could not quote? Those are the ones to watch — that is where bias gets into a process that looks careful.
  • Would your order survive someone else running the same profile? If not, the profile is doing less of the work than you think.

Without a matching tool

Do the ranking by hand for ten people. It takes about forty minutes and is arguably the better version of this exercise, because writing a reason for each person is the whole point.

If you want a machine order to compare against, add a weighted-sum column to the spreadsheet from Module 4. It is a crude keyword matcher, which is useful — it will disagree with you in exactly the ways Lesson 1 describes, and you will be able to see why.

How long to spend

Fifty minutes. The writing is the slow part and it is the point. A ranking you can justify in a sentence per person is one you can defend, improve, and hand to someone else.

Common questions

What is the difference between keyword and semantic matching?

Keyword matching asks whether a token appears in the document — fast, predictable, auditable, and blind to synonyms and to negation, so 'no experience with Kubernetes' matches a Kubernetes search. Semantic matching compares meaning, so paraphrase works, but it drifts toward topic rather than requirement and quietly rewards longer documents.

Should you use AI that learns from your past hires?

Be careful. Unless you feed in genuine performance data, which almost nobody does, the system learns from your past screening decisions — so 'find candidates similar to our best hires' is a specification for repeating the past, including the parts you would not defend if they were written down as criteria.

How do you check whether an AI explanation is real?

Delete the sentence the explanation cites from a test resume and re-run. If the score drops and the explanation changes, it reflects what drove the ranking. If the score drops but the explanation is identical, the explanation is generated separately and is decoration. If the score does not move, something you did not specify is driving the rank.

Check yourself

Four questions. Nothing is recorded anywhere but your own browser — this is for you, not for a score.

  1. 1Why does semantic matching rank a manager of data engineers close to a data engineer?

  2. 2What is the core problem with 'find candidates similar to our best hires'?

  3. 3You delete the sentence an explanation cites, re-run, and the score is unchanged. What does that mean?

  4. 4Why is banding more honest than ranking?