The comparison layer

Talent_Matcher

Talent_Matcher replaces keyword overlap with layered judgement — a sequence of narrow reads that each answer one question and hand the answer on as evidence. It produces the shortlist a senior hiring manager would produce after actually reading the posting and the resume, with the citation behind every verdict.

  • Distributed systems at scaleFound by name
  • Large-scale data engineeringShown by the work
  • Regulated-industry deliveryFound by name
  • Team leadership beyond tenNot found
Illustrative. Every requirement carries its own verdict and the evidence behind it.

Where matching fails

Counting word overlap fails in three predictable ways.

It has been the industry default for two decades, and it is the source of nearly every complaint recruiters, hiring managers and candidates have about automated shortlisting. Underneath all three is one structural problem: a single arithmetic pass discards everything a human reviewer actually weighs.

  1. Vocabulary transfer

    Candidates and agencies paste the posting's language straight into a summary paragraph. A matcher reading for overlap cannot tell that apart from someone who has actually done the work — so the better copy-editor wins.

    What that looks like

    A candidate whose summary echoes the job description ranks above one whose project bullets prove they can do the job.

  2. Career shape hidden by shared vocabulary

    Adjacent specialists share terminology with the target role without sharing its shape. Overlap treats them as equivalent because, on vocabulary alone, they are.

    What that looks like

    A chief architect is rated below a data-migration specialist, because the specialist happens to carry “enterprise architecture” in their title.

  3. Seniority and scope stripped out

    Ten years owning a small workstream and ten years inside a large programme list the same technologies. Nothing in a keyword pass distinguishes them.

    What that looks like

    A nurse practitioner with fifteen years in intensive care is filtered out for writing “ICU” where the posting said “critical care unit”.

How it reads

Four narrow questions, not one broad one.

Asked to judge everything at once, any reader — human or otherwise — does it worse than the same reader asked one bounded question at a time. So the reading happens in stages, each with a single question to answer and no licence to wander into the next.

  1. Structured role understanding

    What does this role actually require?

    Before any candidate is looked at, the posting stops being prose and becomes something that can be tested against. Ambiguity gets resolved here rather than inherited by everything downstream, and nothing in the posting is quietly dropped on the way.

  2. Your review, then a lock

    Have we understood it the way you meant it?

    What we drew from the posting goes back to the recruiter or hiring manager before a single candidate is evaluated. Anything we read wrongly gets corrected, anything missing gets added, anything that does not belong comes out. Then it is locked, and every candidate is measured against that same definition.

  3. Career shape

    Is this the right kind of person at all?

    The shape of a career is read against the shape of the role — what someone has actually been, at what altitude, on what sort of problem, rather than what their title says. A candidate in the wrong lane cannot be rescued by vocabulary overlap, and one clearly in the right lane is not punished for a missing peripheral term.

  4. Requirement by requirement

    For each requirement, what is the evidence?

    Each requirement is settled on its own — found by name, demonstrated by the work described, or not found. Anything other than “not found” must carry the exact resume text that supports it. Because every requirement is judged on its own, a peripheral item never quietly steals attention from a central one.

Running through all four

Questions that are arithmetic are answered by arithmetic.

Judgement is reserved for what is genuinely a matter of judgement — does this piece of work demonstrate this capability? Anything with a defined answer is computed rather than re-decided, so it comes out the same way every time. Whether twelve years satisfies a requirement for ten is not an opinion, and it is not treated as one.

In practice

Three mandates where the keywords pointed the wrong way.

Not searches that returned too little. Searches that returned a confident, plausible, wrong answer — the failure that costs the most, because nobody goes back to check it.

  1. SAP Project Manager

    SAP PM and Scrum Master resumes came back more than 95% similar on keywords. On vocabulary alone the two roles are almost indistinguishable.

    What happened

    The Scrum Master profiles were excluded and the genuine SAP PM candidates surfaced — avoiding a false positive that typically costs a recruiter two to three days of wasted screening.

  2. Senior Java Backend

    A product company's pipeline flooded with profiles tagged “Full-Stack Java”, which matched the posting on every term that mattered to a keyword pass.

    What happened

    Full-stack generalists were separated from engineers with genuinely backend-heavy delivery history. The hiring manager's shortlist-to-interview ratio doubled.

  3. Site Reliability Engineering

    The mandate was written as DevOps but was actually SRE. Keyword search returned an identical shortlist for both, because on paper they are the same role.

    What happened

    Reading the role context — SLO ownership, on-call, error budgets — put true SRE candidates first and filtered out the generalist DevOps profiles that would have been rejected at interview anyway.

What you receive

A shortlist you can defend to anyone who asks.

To the hiring manager who wants to know why. To the candidate who asks what was missing. To an auditor, a year later. Every element of it has visible evidence underneath.

A ranked shortlist
Overall confidence and a contextual-fit read for each candidate, ordered.
A verdict badge per candidate
Strong, partial or weak from the shape layer — with a boundary flag where the verdict genuinely depends on how the role is framed.
A detail view behind every one
Every requirement verdict with the resume text that produced it, and what the read of the candidate rested on.
Gaps and highlights in prose
Written from the actual verdicts and the actual resume — not assembled from templates.
A full audit trail
Enough to reconstruct any position on the list months later, including where a boundary case was flagged for a human read.

The numbers

Twenty-five times the pile. Ten times the minutes.

Prioritisation time scales sub-linearly with the size of the batch, which is what makes reading every resume properly affordable at volume rather than only in principle.

  1. 20+ resumes

    Under 2 minutes

    60–90 minutes by hand, for one traditional recruiter

  2. 100+ resumes

    Under 5 minutes

  3. 500+ resumes

    Under 20 minutes

And what the shortlist is worth

Profile rejection at first client screen
25–40% — typical agency range
Under 10% — CognityHire
Shortlist-to-interview conversion
35–45% — typical agency range
55–65% — CognityHire
Interview-to-offer conversion
15–25% — typical agency range
30–40% — CognityHire
Submissions per shortlist
5–8:1 — typical agency range
2–3:1 — CognityHire

CognityHire delivery metrics from recent enterprise engagements, against typical traditional agency ranges.

Design commitments

The parts that are not negotiable.

No verdict without a citation
A positive verdict that cannot point at the specific resume text supporting it is rejected. That constraint is what makes fabricated evidence impossible and leaves a trail a human can check line by line.
Project evidence outranks prose
Bullets under a named role at a named employer carry more weight than a summary paragraph or a career-highlights list. Where the two disagree, the projects win — which is the exact inverse of how naive matching fails.
Honest ambiguity, not forced answers
A candidate whose verdict would flip under a slightly different framing of the role is flagged as a boundary case and sent for a manual read, rather than being pushed to one side of the line and hidden inside a score.
No taxonomy of its own
The reasoning is domain-invariant; every piece of domain vocabulary comes from your posting and the resume. Nothing is encoded for any particular industry, which is why it behaves the same for a nurse practitioner as for a platform engineer.
Protected characteristics are never read
Family, age, health, religion, race, gender, sexual orientation, political affiliation and criminal history are never extracted, referenced or acted on. Not weighted low — not read at all.
Nothing is retained
Roles and resumes are held for the length of the run and no longer. There is no candidate database accumulating quietly behind this.

Straight answers

What people ask about Talent_Matcher.

Can we correct the system before it evaluates anyone?

Yes, and that step is deliberate. What was drawn from your posting is shown to you first, and you correct anything that is wrong before a single candidate is looked at. Once you lock it, every candidate is measured against that same definition — so two identical candidates cannot get different answers because the criteria drifted between runs.

Can we see why a candidate ranked where they did?

Requirement by requirement. Each one shows whether it was found in the candidate's own words, demonstrated by the work they describe, or not found at all — and anything other than “not found” carries the exact resume text behind it. The shortlist is meant to be defensible to a hiring manager, to the candidate, and to an auditor.

What stops a well-written resume from beating a better candidate?

Project evidence outranks prose. Bullets under a named role at a named employer carry more weight than a summary paragraph, and where a summary claims more than the projects underneath it support, that gap is reported rather than rewarded.

Does it work outside technology hiring?

Yes. The reasoning is domain-invariant — all the vocabulary comes from your posting and the candidate's resume, and no taxonomy is encoded for any particular field. The failure it was built to fix shows up just as often in clinical, legal and finance hiring as in engineering.

Does it read anything it should not?

No. Protected characteristics are never extracted, referenced or acted on — not down-weighted, not read. And nothing is retained after a run: there is no growing store of candidate data behind this.

Ask us who you turned down last time.

Most searches never find out. The candidates filtered out at the keyword stage are the ones nobody reviews, because nobody knows they were there.