The retrospective test: check CandidRanker against a role you already filled

Why a retrospective test beats a demo

A demo shows you a vendor's data on a vendor's role, chosen because it goes well. You learn what the interface looks like and almost nothing about whether the ranking would have helped you.

A retrospective test inverts that. You supply the role, you supply the pile, and you already know which candidate turned out to be the hire. The only open question is whether the ranking would have surfaced them, and how far up.

It also tests the thing that actually decides whether a screening tool survives contact with a hiring team: not the top result, but the middle of the list. Anyone can rank an obvious candidate first. The useful question is whether the people your team spent hours rescuing from page four were ranked where they should have been.

What to bring

• One role you have already closed. Recent enough that you remember the reasoning, ideally one that was hard rather than obvious.
• The job description as it was actually used. Not a cleaned-up version. The messy one is the real input.
• The resumes you received. PDFs, Word documents, scanned pages and photographs are all parsed. Duplicates are detected by content hash before any AI runs, so a folder with repeats does not distort the count.
• Your own answer, written down before you look. Who your team shortlisted, who you hired, and which candidates you nearly missed. Writing it first is what keeps the test honest.

What you get back

Every candidate is scored against that specific job description, and the match score decomposes across eight weighted dimensions: skills, tools, experience, education, certifications, projects, title similarity and semantic similarity. Every dimension is shown with the evidence and reasoning behind it, including the zeros, so a score of nothing on certifications reads as a fact about the resume rather than a verdict on the person.

Mandatory skill coverage is weighted more heavily than optional skills, and missing must-haves are surfaced rather than used to remove anyone. No candidate is ever auto-rejected. Employment gaps appear as neutral facts for discussion and are never deducted from the score.

The ranking report exports as PDF, standalone HTML, CSV and JSON, so you can put the comparison in front of the hiring manager who ran the original process and let them argue with it.

How to read the result

There are three outcomes, and only one of them is boring.

• The ranking matches your team. The engine reads your roles the way your team does with no tuning. That is the boring outcome, and it means you can trust the ordering on a live role.
• The ranking puts your hire lower than you expected. Open the breakdown and find the dimension responsible. It is usually vocabulary: a requirement written one way in the job description and another way on the resume. That is what the workspace terminology catalog and Boolean skill groups with alias matching exist to fix, and fixing it takes minutes.
• The ranking surfaces someone your team passed over. The most valuable outcome and the most uncomfortable one. Read the evidence, decide whether the engine or the original process was right, and note that the answer is now recorded either way.

Scoring is deterministic: the same job description and the same candidate inputs produce the same ranking. So when you change a weight or add an alias and re-run, the difference you see is the change you made and nothing else.

What it costs and what happens next

The MinMaxHR platform is free to use: accounts, workspaces and team invitations cost nothing. Plans price CandidRanker usage volume only. The Free plan covers 100 resumes, 10 job descriptions and 10 hiring decisions, which comfortably covers one retrospective test and usually a second.

There is no auto-renewal anywhere in the pricing. Growth is a one-time payment that activates a 30-day period, with nothing to cancel. So a team can run the retrospective test, run a live role, and stop, without a procurement conversation about a subscription.

If you would rather not do it alone, bring one real requisition to a working session with our team and we will run it with you. You keep the output either way.