How the match scores are calculated
Every candidate the system assesses receives several separate scores. This page explains where each one comes from and which parts are computed rather than judged by an AI panel. It describes the method actually in use, not a simplified version of it.
The system sorts, a person decides
The scores order the work; they never decide it. Every hiring decision, and every rejection, is made by a person at the company. No candidate is rejected automatically.
Skill fit
Computed, not estimated. The system takes the skills the job posting requires and asks, for each one, how much genuine command the candidate's documents actually show.
Each required skill carries its own weight: how important the posting marks it, multiplied by how rare that skill is across the talent pool. A rare must-have therefore counts more than a skill almost everyone lists.
Per skill, the evidence decides the credit. A skill backed by real work experience counts in full. A skill backed by education or a certificate counts 0.7. A skill that is merely listed, with nothing behind it, counts 0.3.
Stated years of experience scale that credit and reach full weight at five years. Less than one year counts as a claim, not command. If someone states more years for a single skill than their whole career spans — more than twice over — the claim is treated as unverified and drops back to 0.3.
Certificates multiply onto an existing match. They can lift a real skill overlap, but they can never create one out of nothing.
Only when a job posting lists no skills at all does this number fall back to the AI's context judgment. The dossier then marks it as an AI value instead of a computed one.
Context
A judgment by the AI panel, not a calculation. One of three seats — the expertise seat — answers a single question: how much demonstrated, role-relevant experience and depth does this person have in the skills the role requires?
That seat judges nothing else. Seniority, expectations and the candidate's situation belong to other seats. The dossier shows the panel's evidence and its reservations next to the number.
Personal fit
Also an AI judgment, made by the leadership seat of the panel: how well seniority, ownership, career trajectory and the candidate's own expectations match the level this role actually offers.
This value is deliberately not shown on a candidate. It feeds the overall value described below and nothing else — a single number about how well a person fits invites being read as a verdict on the person.
Constraints
Computed from four dimensions that either match or do not: salary, earliest start date, languages and location. They carry different weights — salary 0.5, start date 0.25, languages 0.25, location 0.25.
Salary: if the posted range and the candidate's expectation overlap at all, the dimension is fully met. Otherwise the score falls with the gap and reaches zero at 15 percent of the posting's midpoint, or 5,000 in the posting's currency, whichever is larger. Different currencies on the two sides mean the dimension carries no comparable data.
Earliest start date: fully met when the candidate is available by the required date, then falling to zero over 90 days of delay.
Languages: for each language the posting requires, meeting or exceeding the level counts fully, a lower level counts proportionally, and a language the candidate does not list counts zero. The result is the average across all required languages. Two cases make the dimension carry no data at all rather than a low score: a posting that states no language requirement, and a profile that lists no languages whatsoever — the latter counts as unknown, not as a failure.
Location: within 25 km counts fully, then falls to 0.6 at 150 km and 0.1 at 1,000 km. A candidate in another country without usable coordinates counts 0.2. A posting that explicitly recruits Europe-wide or internationally raises the floor to 0.5 or 0.8. A remote role has no location dimension at all.
This is the part worth understanding: a dimension with no data on either side drops out. It is not counted as zero and not counted as full — the remaining weights are simply renormalised. Nothing unknown about a candidate can cost them points, and nothing unknown can earn them any.
One consequence is visible daily: on profiles that came from an uploaded CV, this value is frequently absent altogether. A CV states no salary expectation and no earliest start date, so at most language and location carry data. And when a fully met language requirement is the only dimension left, the system reports no constraint value at all rather than a perfect one — otherwise merely having listed a language would buy a full score.
Overall value
Composed in code from the values above, never produced by the AI. Skill fit counts 40 percent, context 35 percent and personal fit 25 percent.
When a constraint value exists, that blend is scaled to 85 percent and the constraint value takes the remaining 15 percent. Whenever no constraint value exists — because no dimension carried data, or because the language-only case described above suppressed it — the three-way blend applies unchanged.
The 40/35/25 split was calibrated against a labelled reference set and confirmed in a blind recruiter test. The 15 percent given to constraints is not calibrated: it is a pilot value, recorded as such in the code, still awaiting the same treatment against the reference set. Both are versioned and re-examined whenever that set changes.
Recruiters never see this combined number: neither the triage cockpit nor the candidate dossier renders it, and the same holds for personal fit. They show the individual dimensions instead, each marked as computed or AI, so that a decision rests on what a number is made of rather than on the number alone. There is one place where the combined value does appear — a talent who receives an invitation sees it on that invitation, as a match percentage.
One final, capped step is added when the posting carries an assessment and the candidate has completed it: the composed value is then multiplied by at most 1.05. The full five percent is reached only by a flawless questionnaire; below that the boost falls linearly with the result achieved. Its source is the assessment value stored on the case — the share of preferred answers. The boost acts on this combined value alone and on none of the individual dimensions. It can therefore change the order among people who are all but level, but it can move nobody into another triage group and overrides no eligibility decision — that one is taken two steps earlier. Anyone without an assessment, with one still open, or excused from it is computed unchanged: a missing assessment costs nothing.
Apprenticeships: a standard of their own
Apprenticeship postings are judged under their own standard. It assesses requirements, aptitude and reliability, not work experience: nobody needs to bring that to an apprenticeship.
Requirements weigh 40 % and are computed: language level, a school qualification by the start of training and required certificates, averaged with the formal review's assessment. Aptitude and reliability weigh 30 % each and come from one AI seat each.
School marks in the subjects defined for the occupation lift the requirements by at most 10 points. They never lower a score, and missing marks are no disadvantage.
Age, origin, nationality, gender, health and family are never used, inferred or commented on. If an assessment names one of them, its text is withheld, its score discarded and the recommendation set to review.
How the groups A, B and C are formed
The groups are a work order, not a rating of people. They say where to start, not who is best.
A recruiter who moves someone into another group overrides everything below. The derivation is kept and named alongside the move, not erased. Where no such move exists and the AI panel reached a binding verdict, that verdict places the candidate: advance goes to A, review to B, reject to C.
Without a panel verdict the displayed dimensions decide. Any dimension below 40 places the candidate in C. All dimensions above 60 place them in A. Everything else lands in B. Whether a panel verdict or the dimensions place them: A receives at most seven people automatically, and everyone after them follows in B. If a run places nobody in A automatically, up to seven people with the highest overall value move up from B. People moved by hand or already decided do not count here.
Dimensions that carry no data take no part in that test — an absent value can push nobody up and nobody down.
A fourth pill marked with a question mark collects profiles that a run handed to the panel without getting a judgment back, for example after an aborted run or an unreadable CV. It appears only when it is not empty.
Limits worth knowing
The scores pre-sort documents; they do not measure a person. They can only see what the documents contain — a CV that omits something is silent about it, not negative.
The two judging seats are language models. Their assessments are stored with the evidence and the reservations behind them so they can be checked rather than taken on trust.
Questions about an individual assessment go to the company that posted the role. It holds the dossier and makes the decision.