Read an Application in Order of What You Can Check

What should you read first, and why that order?

Start with the facts a third party could confirm, because those are the only parts of the document whose cost of production has not fallen. Employment dates and continuity. Employer names. Location and work authorization. Licences and registrations. Then specific claims attached to a named place, and only then the prose. The order is doing the work, since a reader who stops after twenty seconds has decided on whatever occupied those twenty seconds.

The inherited order does the opposite. Every widely ranking answer to this question is written for the applicant: the six-second scan, the F-pattern, the heat map with eyes on name, title, company and the first bullets. Employers absorbed it without examining it, and it front-loads the headline, the summary and the achievement bullets, which are the three parts of a document that are now free to generate.

Making a claim machine-checkable appears to move what people write. Verifying every publication listed by applicants to one orthopaedic surgery residency programme, researchers found 13 misrepresented citations out of 1,100, or 1.18%, against 18.0% in the same programme's 1999 study and 20.6% in its 2007 study, and they attribute the fall to the application form adding the PubMed Identifier ¹. Hold the limits: one programme, one specialty, a uniquely checkable kind of claim, a narrow definition of misrepresentation, and the PubMed explanation is the authors' hypothesis rather than a tested causal result. Other specialties measured in the same period report much higher rates, so this is not a general honesty rate for applications.

The mechanism transfers even where the number does not. A field that resolves against an external index behaves differently from a field that does not, which is why the checkable spine deserves the first slot and why the identifier belongs on the form.

Which claims on an application can actually be checked?

The ones that resolve against something outside the document. A registration number against a regulator's list. A date range against an employment verification. A named client, system or venue against the public record. A named publication against an index. Everything else on the page is a description of the person written by the person, and its accuracy is unknowable at the screening stage no matter how long you read.

Rank the page by that test and it sorts into three tiers:

• Checkable now, for free. Dates, employers, licences, registrations, named places, anything with an identifier. Read these first, in the first fifteen seconds.
• Specific and falsifiable, checkable later. "Migrated a named system at a named company in a named year." You cannot confirm it at the screen, but you can ask about it, and specificity carries real cost to invent in a way that generality does not.
• Not checkable at all. Summary statements, adjectives, achievement bullets with no place attached, and any number without a named context. Read last, weight at zero.

"Improved efficiency 40%" belongs in the third tier, not the second. The advice to hunt for percentages and dollar figures inverts the value now: a specific number tied to a named place is worth reading, while a generic one is the easiest line on the page to produce. Where the claim is about tooling rather than results, verifying an AI-proficiency claim without taking their word for it is the same sorting applied to a subject with no registry behind it.

Experience thresholds sit awkwardly in the first tier and deserve a warning label. Dates are checkable, but what they carry is thinner than the screen assumes: a meta-analysis of 81 independent samples put prehire work experience at a corrected correlation of .06 with later job performance, and about .07 for experience with relevant tasks, jobs or occupations ². Check the dates for continuity and scope. Do not read a year count as a quality measure.

Why polish stopped being a signal in either direction

Because you cannot tell from the page where it came from. Untrained evaluators asked to separate model-written text from human writing performed at random chance, and three quick training methods lifted them only to about 55%, inconsistently across the three text domains tested ³. That was 2021-era output judged by crowdworkers. Newer output is unlikely to be easier, and whether an experienced reader in their own field does better was not tested.

This cuts both ways, and both directions are being got wrong in practice. Treating a polished application as evidence of care rewards whoever had the better drafting tool. Treating it as evidence of generation punishes people who write well, people writing in a second language, and anyone who used a tool the way a professional would. The second error is the more expensive one, because it produces a rejection with a reason nobody can state.

Corpus-level estimates do not rescue individual judgement either. Applying a maximum-likelihood method to scientific peer reviews, researchers estimated that between 6.5% and 16.9% of the review text submitted to four 2023 and 2024 conferences could have been substantially modified by a language model, and the same paper says these corpus-level trends may be too subtle to detect at the individual level ⁴. Knowing roughly how much of a pile involved a model still tells you nothing about the document in front of you.

Specifics are the only part of free text left worth reading. Polish is not a criterion at this stage, in either direction, and authorship is not worth screening attention at all. That is where what to screen on when every resume looks perfect picks up, and it is also the reason a portfolio needs a conversation rather than a closer read, as in working out who actually built the portfolio project.

Set up the review screen so the checkable fields come first

Change where the fields sit rather than trying to change how people read. Where there is an applicant tracking system, screening happens on a review screen somebody configured once, and whatever that screen shows above the fold gets read first regardless of any policy. Reordering it is usually a ten-minute configuration change and the highest-return move available at this stage.

A sensible arrangement, top to bottom:

1. Employment history with dates, displayed as a continuous timeline, so gaps and overlaps are visible without arithmetic. 2. Licences, registrations and identifiers, as their own fields on the application form rather than buried in a document. A field asking for a registration number gets a registration number. 3. Location and work authorization, as structured answers. 4. One free-text answer to a role-specific question, which is where the specifics live. 5. The attached document, last, for anyone who reaches the expensive read.

Employers who screen this way are not a small population. In the Harvard Business School and Accenture survey of 2,275 executives, 63% reported using a recruiting management system, rising to 69% above a thousand workers, and more than 90% of those employers used it to initially filter or rank candidates, specifically 94% for middle-skills and 92% for high-skills roles ⁵. Read that carefully, because it is widely misquoted: it is criterion matching configured by a person, the base is the 63% who reported having such a system, and the fieldwork was early 2020, so it says nothing about model-based screening. What it does establish is that for most employers in that survey, the first pass happens on the review screen and under its rules.

Two things to do in the same session. Add the identifier fields to the application form, since a field is what makes a claim checkable and a paragraph is what makes it a story. Then sample what the new order rejects, because a reordered screen is a changed screen, and the only way to find out whether it is cutting the right people is to look at the ones it cut.

Reading in order of what you can check tells you who is real and plausibly in scope. It does not tell you who is good, and no ordering of a document ever will, which is the argument for moving that judgment to a stage where somebody produces work in front of you.

See a sample report >