The Search Brief Acceptance Standard: Source, Revise, or Send Back

This is the acceptance test I run on a role brief before anyone spends a day sourcing against it or hands it to an AI sourcing agent. It is for in-house recruiters, sourcers, talent leaders, and founders who do their own hiring. It gives you a fixed set of criteria, a 10-profile two-sourcer agreement test, and three verdicts (Source, Revise, or Send Back), defined tightly enough that two graders reach the same one and every failure names the field and the question that fixes it.

What makes a search brief "accepted"?

A search brief is accepted when every must-have is written as evidence a sourcer can see on a public profile or is explicitly deferred to the screen, the trade-offs are ranked on paper, every proxy is justified or struck, and two sourcers grading the same 10 profiles blind agree on at least 8 of them. Anything short of that is Revise or Send Back, and which one depends on who owns the field that failed.

Four terms carry the standard:

• Search brief: the written spec a sourcer, human or agent, works from to decide whether a profile is a fit.
• Must-have: a requirement the hiring manager would reject on even if everything else about the candidate were strong.
• Proxy: a stand-in for capability, such as years of experience, a degree, or a named former employer.
• Boundary profile: a profile that clearly meets some criteria and only arguably meets others, so a vague brief produces split grades on it.

Published intake guides stop at a list of questions and a template. They tell you what to ask, not when the answers are good enough to source from. This standard draws that line. It runs before any sourcing, which makes it a different job from calibrating after the first profiles go out. Calibration tunes a spec against live feedback. This gate fails a brief before the hours are spent.

The three verdicts

Which hiring-manager rejections can a better brief prevent?

Most merit-based rejections fall into categories a brief field already covers: seniority, scope, domain, and skills. If those fields are vague, the hiring manager rejects against criteria the sourcer never saw.

One recruiting platform published a study of more than 500,000 merit-based rejection decisions across 2,000+ organizations and 20,000+ users. Five of its top reasons map directly to brief fields. The mapping in the right-hand column is my desk analysis, not part of the study.

Three things follow.

1. Rejections hit a band, not a floor. Too much experience is rejected nearly as often as too little. A brief that states only a minimum leaves half the seniority decision unwritten. Every seniority criterion needs an upper bound too.
2. Rejections are fast. The same study found 64.8% of merit rejections are decided within an hour of a candidate entering the pipeline. That speed means the hiring manager is matching against a picture already in their head. If that picture is not on paper, the sourcer cannot match it.
3. Some rejections are outside the brief's reach. The study excluded "never responded to outreach" (28.7% of all reasoned rejections) and "job closed or archived" (21.3%) from its merit analysis. No brief fixes those.

Know the limits of this evidence. The reasons were logged by recruiters, not only by hiring managers. Nobody has published what share of these rejections a better brief actually prevented. Treat the table as a map of where vague briefs hurt, not as a promise of how much a good one saves.

The acceptance criteria, field by field

The standard has seven criteria. Each one below says what passing looks like, what it proves, what it looks like when it lies, and who owns the fix.

1. Scope and domain are written.
   • Pass: one sentence on what this person will own, and one naming the industry or problem space they must come from, with the reason.
   • Proves: sourcers will search the same kind of role in the same kind of company.
   • Lies when: the domain is a list of logos. That is a named-employer proxy dressed as a domain (see criterion 5).
   • Owner: hiring manager.
2. The seniority band has two bounds, stated as capability.
   • Pass: a lower bound ("has already owned X end to end") and an upper bound ("has not yet run Y, or would be bored by Z").
   • Proves: both sourcers will reject the same overqualified and underqualified profiles.
   • Lies when: the band is a years range. Years let two sourcers agree on a number while disagreeing about the person.
   • Owner: hiring manager.
3. No more than four must-haves, each surviving the test.
   • Pass: four or fewer, and for each the hiring manager answered "no" to "would you hire someone who had everything else but not this?" One widely used intake template caps must-haves at four and attributes this test to LinkedIn's intake guidance. I have not verified that attribution, but the test stands on its own.
   • Proves: the must-haves are actual reject criteria, not a wish list.
   • Lies when: the hiring manager agrees to four but keeps a fifth in their head. The agreement test usually surfaces it.
   • Owner: hiring manager.
4. Every criterion is tagged for visibility.
   • Pass: each criterion is marked Profile (a sourcer can see the evidence on a public profile) or Screen (it can only be confirmed in conversation), and no Screen item is used as a sourcing filter.
   • Proves: sourcers grade only what they can see.
   • Lies when: a Screen item such as compensation, availability, or work authorization stays in the must-haves. Sourcers then guess, and they guess differently.
   • Owner: recruiter.
5. Every proxy is justified or struck.
   • Pass: each years, degree, or named-employer requirement is rewritten as the capability it stands for, or carries a written reason it is job-related.
   • Proves: the brief screens for ability, and you have a record of why.
   • Lies when: the justification just repeats the proxy ("five years because the role is senior").
   • Owner: hiring manager, because only they can say why the proxy matters.
6. Nice-to-haves are ranked.
   • Pass: a numbered list in the hiring manager's order.
   • Proves: when two boundary candidates each have different nice-to-haves, both sourcers prefer the same one.
   • Lies when: the list is numbered but the hiring manager says "they're all about equal". That is an unranked list with numbers on it.
   • Owner: hiring manager.
7. The title list is written and tested.
   • Pass: included and excluded title strings are listed, and each included string has been run alone with 25 results read.
   • Proves: both sourcers start from the same pool.
   • Lies when: a string matches out-of-field profiles nobody looked at (see the next section).
   • Owner: recruiter.

The 10-profile test then checks whether these criteria actually produce agreement. It has its own section below.

Criteria 3 and 4 interact, and the interaction decides where each requirement lives in the brief.

Use this skeleton as the brief itself. Every field maps to a criterion, so grading it is a matter of reading down the page.

ROLE BRIEF: [role title]

1. SCOPE (what this person owns and delivers):
1. DOMAIN (industry or problem space, and why it matters):

2. SENIORITY BAND
   Lower bound (capability they must already show):
   Upper bound (what would make them overqualified, and why):

3. MUST-HAVES (max 4; each passes "everything else but not this?")
   a. [requirement] | Evidence a sourcer sees: [ ] | Visibility: Profile / Screen
   b.
   c.
   d.

5. PROXIES
   [years / degree / named employer] -> Capability it stands for: [ ]
   Kept because (job-related reason): [ ]   OR   STRUCK

6. RANKED NICE-TO-HAVES
   1.
   2.
   3.

7. TITLE STRINGS
   Include (each run alone, 25 results read):
   Exclude (tested, out of field):

EXAMPLE PROFILES (3 clear fits, with one line on why each fits):

SCREEN-STAGE QUESTIONS (comp, availability, work authorization, location):

AGREEMENT TEST: [x]/10 agree | Must-have splits: [n] | Boundary profiles: [n]
VERDICT: Source / Revise / Send Back
Failed field:            Fixing question:
Graded by:               Date:

Why is the title list the first place two sourcers disagree?

Because small changes to the title list swing the pool by thousands of profiles, and an ambiguous title string pulls in people from outside recruiting. Two sourcers who expand titles differently are searching different populations before they read a single profile.

I measured this on sourcer roles themselves. In Refolk's index of professional profiles, the counts move like this:

The four titles were Sourcer, Talent Sourcer, Technical Sourcer, and Sourcing Specialist. The US pool is 11.9 times the UK pool on the same title set.

The size is only half of it. I read the top 25 profiles for each query:

In the UK sample, 12 of 25 profiles carried "Sourcing Specialist", and the same sample included a Procurement Specialist, a Billing Administrator, and a Professional Services Category Lead. "Sourcing" means purchasing in procurement, so the string pulls in buyers. In the narrowed US sample, 22 of 25 were titled Technical or Talent Sourcer, which is a much cleaner pool, but it also drops every sourcer who uses a different title.

So criterion 7 is not paperwork. The test is simple: run each title string alone, read 25 results, and write down which strings go in and which go out. In Refolk you can describe the pool in plain English, including what to exclude, and read the top results in minutes instead of building boolean strings for each variant.

The 10-profile two-sourcer agreement test

The test is a smoke test: two sourcers grade the same 10 real profiles blind against the brief, and the brief passes only if they agree on at least 8 and split on no must-have. It cannot prove a brief is good. It reliably shows where a brief is vague, because every disagreement traces to a field.

Build the sample

• 10 real profiles, chosen by someone who will not grade them.
• At least 3 boundary profiles. These are the ones that test the brief. A vague seniority band or an unranked trade-off only shows up on a profile that sits near the edge.
• Clear fits and clear misses for the rest, so graders are not primed to see every profile as borderline.

Sourcing the boundary cases is the slow part. Describing the edge in plain English in Refolk ("platform engineers who led a migration but never managed a team") is a quick way to pull real profiles that sit exactly on the line you want to test.

Grade

Each sourcer, working alone and without seeing the other's sheet:

1. Marks each profile Fit or No Fit.
2. Marks each must-have pass or fail for that profile.
3. For every No Fit, names the criterion that drove it.

Score

• Percent agreement = profiles with the same Fit or No Fit call, out of 10. Each disagreement costs 10 points.
• Must-have splits = any profile where the sourcers disagree on whether a single must-have is met, even if their overall calls match.
• Pass = 8 or more of 10 agree, zero must-have splits, and the sample held at least 3 boundary profiles.

Why 8 of 10, and why not kappa

Cohen's kappa is the standard statistic for agreement between two raters beyond chance. On Cohen's own scale, 0.41 to 0.60 is moderate and 0.61 to 0.80 substantial. McHugh's review of the statistic argues that reading is too lenient, that any kappa below 0.60 shows inadequate agreement, and that many texts set 80% raw agreement as the minimum acceptable. My 8 of 10 threshold borrows that 80% floor. No published threshold exists for a 10-item test, so treat 8 of 10 as a policy choice, not a statistical result.

The zero-split rule on must-haves is what makes 10 profiles useful despite the small sample. One split on a must-have means two careful people read the same requirement two ways. That is a brief defect regardless of the overall score.

The procedure: from intake to verdict

Run these eight steps in order. Steps 1 to 5 produce the brief, steps 6 and 7 test it, and step 8 records the verdict.

Sources disagree on when calibration belongs. One published method has recruiters go through found resumes line by line inside the intake meeting. Another calibrates after the first three to five profiles are sent. One intake agent calibrates during the conversation itself by surfacing up to five matching CRM candidates at a time. This standard runs between intake and sourcing because that is the last point where a failure costs nothing. It does not replace calibrating on live results later.

Tracing a disagreement to a field

Step 7 only pays off if each disagreement turns into a named field and a question. This is the loop:

These are the fixing questions I use for each hiring-manager-owned field:

When the verdict is Send Back, keep the message to one field and one question. A list of six concerns gets a vague answer to all six.

Subject: [Role] brief: one decision needed before sourcing starts

Hi [name],

Before starting the search, I had two sourcers grade the same 10 profiles against the brief, independently. They agreed on [x] of 10. The disagreements came back to one field:

Field: [field name]
What happened: [Sourcer A read it as ___; Sourcer B read it as ___, on profiles ___]
The question that fixes it: [one question]

Once I have your answer, I will update the brief, re-run the same 10 profiles, and start sourcing.

Thanks,
[name]

How the standard goes wrong

The standard fails in six predictable ways. Each one produces a verdict that looks right and is not, so each comes with a check.

1. Ambiguous title strings.
   • What happens: a string like "Sourcing Specialist" matches procurement profiles as well as recruiting ones. In Refolk's UK sample, the broad query surfaced at least four non-recruiting titles in 25.
   • False positive: buyers and category leads counted as sourcers, inflating the pool and distorting the sample.
   • Check: run each title string alone and read 25 results before it goes in the brief.
2. Kappa on 10 profiles.
   • What happens: someone computes kappa on the test and reports it as a reliability score. Guidance on small samples warns that even a kappa from 20 items can swing wildly.
   • False positive: a brief certified on a number that would change with one different profile.
   • Check: use percent agreement only, and enforce zero must-have splits.
3. High agreement on an unbalanced sample.
   • What happens: a sample of nine clear misses and one fit gets 10 of 10 agreement because nothing in it was hard. Skewed prevalence also distorts kappa, which is one more reason not to use it.
   • False positive: a vague brief passes because the sample never tested its vague fields.
   • Check: require at least three boundary profiles, and have someone who will not grade choose the sample.
4. A years band used as the seniority criterion.
   • What happens: two sourcers agree on the number and reject capable people with thin CVs. In one dataset of 16,308 hiring decisions, candidates rejected for insufficient experience scored 3.51 on a practical task, against 2.40 for candidates who failed the test.
   • False negative: strong candidates screened out by tenure.
   • Check: for every years band, name the capability it stands in for and grade that.
5. A must-have the profile cannot show.
   • What happens: compensation, availability, or work authorization sits among the must-haves. Sourcers guess, they guess differently, and agreement fails for reasons that have nothing to do with fit.
   • False negative: a precise brief sent back because of one misplaced field.
   • Check: move it to the screen-stage questions and re-grade.
6. Vendor metrics taken as proof the brief is good.
   • What happens: agent performance figures are quoted as if they validate your brief. One agent's reported reduction in profiles reviewed has appeared as 62% in launch coverage and 81% on the current product page. Another vendor reports 83% of recommended candidates accepted into pipelines. None of these discloses the intake inputs behind it.
   • False positive: confidence borrowed from numbers that say nothing about this brief.
   • Check: measure your own slate acceptance against your own baseline.

There is one more limit, and it is the standard's own. Passing the test means two sourcers read the brief the same way. It does not mean the brief matches what the hiring manager actually wants. A hiring manager who agrees to four must-haves and quietly holds a fifth will pass this gate and reject the slate anyway. That is what post-launch calibration is for.

Briefing an AI sourcing agent, and the record the brief leaves

An AI sourcing agent cannot return a shortlist sharper than its brief, so the accepted brief is the ceiling on what it can find. The same document also becomes your written record that each sourcing criterion is job-related.

What the agent needs that a human infers

Agents now take two inputs that a vague brief leaves out:

• Example profiles. LinkedIn's hiring agent lets recruiters paste the URLs of ideal candidates so it calibrates on real people instead of adjectives. The three clear fits in the skeleton's example field serve this purpose, and they should come from the test sample.
• Hard non-negotiables. In at least one sourcing agent, non-negotiables sit outside the learning loop, so feedback cannot wear them down. The must-haves that survived step 2 are your non-negotiables. If you never wrote them down, the agent infers them from feedback.

Some agents confirm Screen fields through outreach rather than the profile. LinkedIn's agent confirms location preferences, work authorization, availability, and answers to recruiter-defined questions automatically over InMail. That is one more reason to tag visibility: Screen items become outreach questions, not filters. How each field maps to evidence across other agents is not publicly documented, so check what your agent can actually read before assuming it grades a field.

The record regulators expect

• EU. The AI Act's Annex III, point 4(a), lists systems used "to place targeted job advertisements, analyse and filter job applications, and evaluate candidates" as high-risk. The Article 6(3) exemption rarely helps, because an Annex III system that profiles people is always high-risk. Under the Commission's draft guidelines, generating ad creative from human-supplied criteria is not high-risk. The application date for these obligations has already been moved once by amending regulation, so check the current date in the Official Journal rather than relying on one you read somewhere.
• US. Under Title VII, a selection tool with adverse impact must be "job related for the position in question and consistent with business necessity." Adverse impact is commonly tested with the four-fifths rule: a selection rate for any protected group below 80% of the rate for the highest-selected group. The EEOC has moved to rescind its Uniform Guidelines on Employee Selection Procedures and has removed its AI technical-assistance documents. Rescission removes the roadmap many employers relied on, not the burden itself.

Specific regulator guidance on years, degrees, or named employers as brief criteria is not established in the sources I reviewed. The practical point holds anyway: criterion 5 produces, for every proxy, either a written job-related reason or a strike-through. Filed with the verdict, that is the record.

Keeping the standard current

The standard only stays useful if you check it against outcomes. Run the checklist on every brief, then review the thresholds and title tests on a fixed cadence.

• Track slate acceptance against a baseline. Record what share of sourced profiles the hiring manager accepts, for briefs that passed this gate and for briefs that skipped it. That comparison is the only evidence that will tell you whether 8 of 10 is the right bar for your team.
• Re-run title tests. Title vocabularies drift. Re-run each title string you rely on and read 25 results whenever a role reopens, not just the first time.
• Log every split. Keep a running tally of which field caused each disagreement. If seniority upper bounds cause most of your Send Backs, add that question to the intake template itself.
• Re-check the legal mechanism. Before relying on any regulatory date or guidance document, confirm it is still current at the source.