Every Requirement Needs a Stage That Tests It

What makes a requirement testable?

A requirement is testable when the evidence for it has a name and a place: what a person would do or show that satisfies it, and the stage where they will do or show it. Attributes fail that test. *Strong communicator* names nothing observable, so two interviewers can rate it opposite ways and both be right. *Explains a technical decision to somebody who disagrees, in round two* gives the interviewer something to listen for.

Take the attribute, ask what the person is actually going to do with it in the first quarter, and write that down instead.

• Strong communicator becomes *explains a decision to a colleague who disagrees with it*, tested in round two by handing over a real disagreement from last quarter.
• Detail-oriented becomes *finds the error in a document that looks finished*, tested with a short exercise built from a document that had an error in it.
• Self-starter becomes *picks what to work on when the brief is ambiguous, and says why*, tested in the screen with a question about one specific week.
• Five or more years of experience becomes whatever you thought the five years would buy, because the years themselves carry very little. A meta-analysis of 81 independent samples puts the corrected correlation between prehire work experience and later job performance at .06, and experience with relevant tasks, jobs or occupations does no better, at roughly .07 ¹.

The experience line is the one people fight hardest to keep, so read the finding carefully. It covers experience measured at hire, which is exactly what a resume screen reads, and it says nothing about licensure or a legally required minimum. Within that scope, the crude count you are screening on is not carrying the signal you think it is.

Relevance is what carries signal. In the job knowledge meta-analysis behind the 2022 revision of the selection literature, all 164 studies together averaged an observed validity of .22, while the 59 studies using knowledge tests built for the job in question averaged .31, rising to .40 once corrected for unreliable performance ratings ². That is a comparison between subsets rather than a controlled trial, and job knowledge tests assume candidates who already hold the knowledge. Within that one meta-analysis, the job-specific subset predicted better than the rest, which is the reason to write a requirement around the work itself.

Write the list as two columns before the posting goes out

Two columns, one page, before anyone opens the posting template: the requirement on the left, the stage that will produce evidence for it on the right. Fill the right column out loud with whoever is going to run that stage, because a stage nobody has agreed to run is not a stage. Then read down the right column and count how many times each name appears.

Three patterns show up the first time anyone does this. One stage carries eight requirements and the rest carry none, which means the loop has one real stage and three social ones. Two requirements map to the same stage and contradict each other, usually speed against thoroughness, and the interviewer has been silently choosing between them for months. And several lines map to nothing at all, which is the point of the exercise.

The right column also settles arguments the left column cannot. Whether an AI-skills line belongs in the posting stops being a question of principle once you have to name the stage that will check it, which is also most of the work in writing an AI-skills requirement that isn't legally vague. A requirement with a stage behind it is a commitment. A requirement without one is a preference wearing a requirement's clothes.

Why does an untested requirement cost more than it used to?

Because it comes back to you confirmed. A requirement nobody checked used to be decoration: it sat in the posting, nobody asked about it, and the loop carried on without it. Now the posting is the brief that applications get assembled against, so every line in it is answered on the way in, and the process has no way to tell a confirmation apart from evidence.

The gap between what a posting says and what a process does has been measured directly, in the one case where an employer's stated requirement changed en masse. Matching 11,332 roles at large firms that dropped a degree requirement against the career histories of roughly 65 million US workers, the Burning Glass Institute and Harvard Business School found those firms raised the share of hires without a bachelor's degree into those roles by about 3.5 percentage points. Because only 3.6% of roles dropped a requirement, the net effect across the labour market was 0.14 percentage points, which the authors put at roughly 97,000 workers a year and fewer than 1 in 700 hires ³.

The interesting half is who moved. Nearly all the real change came from 37% of firms, which raised their share of hires without a degree in the analysed roles by nearly 20% in relative terms; about 45% changed the posting with no meaningful difference in hiring behaviour, and about a fifth made gains that did not stick ⁴. Education there is inferred from career histories rather than verified, and the population is only firms posting more than 500 job ads a year, so these are estimates about large employers. Within that population, editing requirement text changed hiring only where the rest of the process changed with it, which is the argument for dropping the degree line only in the same change that replaces it.

The right-hand column exists for exactly that reason. Without a stage behind it, the edit stops at the document, and the document is the surface applications are now written toward, which is also why keyword screening on that document stopped working.

Cut the list until every line has a stage

Delete every line whose right column is empty, then check whether the stages you named can carry what you assigned them. Most loops have one stage doing nearly all the filtering and several that mostly ratify it, so a requirement parked in a late round is a requirement almost nobody is testing. Either move it earlier or cut it.

Across more than 54 million applications and 93,000 jobs in one ATS vendor's customer base, recruiter screens passed roughly 35% of the candidates who reached them, while post-onsite stages converted at 95% and offer stages at 81% ⁵. Those are passthrough rates among candidates who got that far, not end-to-end rates, and the customer base skews to venture-backed technology companies, so treat the levels as illustrative. The pattern is the useful part: by the time a candidate reaches the last stage, the decision is effectively already made, so a requirement you assigned to the final round gets ratified on the way past.

On Monday, take the open req and do four things in this order.

1. List the requirements as they are actually written, including the ones copied from the last version of this role. 2. Beside each, write the stage and the name of the person who will run it. A round number is not an owner. 3. Strike every line with a blank beside it. Read the struck lines aloud to the hiring manager; the ones they fight for are the real requirements, and now they need a stage. 4. Count what is left. If more than five survive, go back to the stages: a loop that cannot cover five requirements will cover four and guess at the rest.

What this does not settle is where the bar sits on the lines that survive, which is a separate and harder piece of work covered in setting a defensible bar for good enough in a specific role. Two columns tell you which requirements the process can speak to. They do not tell you how much is enough. But a bar on a requirement nobody tests is not a bar at all, so the columns come first.

See what gets scored >