How candidates are scored

• The assessment measures how well a CV and screening answers match the requirements of one specific vacancy — not the person, their character or their “potential in general”.
• Only information the candidate provided is used: the CV file and the answers in the application form. Social networks, search engines and other outside sources are not used.
• Candidates for different vacancies are never compared with each other. The same person can score differently for different vacancies — that is expected, because the requirements differ.
• The assessment makes no decisions. Rejecting, inviting or any other decision is made by the employer.

1. The candidate confirms their email address with a code. Nothing is processed by AI before that.
2. Text is extracted from the PDF. A scan without a text layer is not assessed — the candidate is told immediately, on the application page, and can upload another file.
3. An AI model breaks the CV down into fields: positions, companies, duration, skills, achievements, education. The model is explicitly forbidden to invent data: what is not in the CV is not in the result.
4. Total experience is determined by the software under the rules in section 6, not by the model.
5. Information that identifies the person or reveals protected characteristics is removed (section 7.1).
6. First assessment: the primary AI model gives a score from 0 to 100 for each of five criteria, with a short justification.
7. The software computes the final score and category with the formula in section 4.
8. If the score is close to a category boundary, a second model from a different developer independently re-assesses the application (section 5).
9. Only when the employer asks, and only for the top 3–5 applications of one vacancy, the second model compares the finalists with each other, using the same de-identified data and the same rules.

Industry relevance — 20%

How closely the industry and domain of past work match the vacancy's industry.

Measurable results — 15%

Whether the CV states results in numbers: growth, volumes, deadlines. No numbers means 0.

Company-scale match — 15%

Whether the candidate has worked in the kind of company the vacancy names: start-up or large company.

Progression and motivation — 15%

Growth in responsibility across the career and how specific the answers to the vacancy's questions are. Employment gaps and short tenures do not lower this score by themselves.

Fit with the vacancy requirements — 35%

Whether the candidate has the required skills listed in the vacancy. Missing required skills cap this at 49; not meeting the minimum experience caps it at 39.

The weights are the same for every candidate and every vacancy. The model cannot change them: they are set in the software, not in the text sent to the model.

The final score is the weighted sum of the five criterion scores: each score is multiplied by its weight, the products are added up and the result is rounded to one decimal place. The software does the calculation, not the model: any total the model suggests is not used and is stored only to detect discrepancies.

The category depends on the score alone: 75 or more — “High priority”; 50 or more but below 75 — “Standard”; below 50 — “Low”.

Worked example:

Industry relevance

80 × 0.2 = 16

Measurable results

60 × 0.15 = 9

Company-scale match

70 × 0.15 = 10.5

Progression and motivation

50 × 0.15 = 7.5

Fit with the vacancy requirements

90 × 0.35 = 31.5

Final score

74.5 — category “Standard”, inside the review band: the application goes to an independent re-assessment

If the model returns fewer than five criterion scores, the software neither fills in a zero nor re-weights: the model's own total is used, the category is still derived from the thresholds, and the case is flagged. If the model's answer is unusable altogether, the application is marked as an error and assessed again — a software failure never turns into a low score for a person.

A score within 7 points of a category boundary — from 43 to 57 and from 68 to 82 — is treated as borderline: in that range a small model error changes the candidate's category.

Such an application is assessed again by a second model from a different developer. It sees the first assessment for reference but must assess the candidate on its own and actively look for anything the first model missed. The second model's result replaces the first. If the second model is unavailable, the first assessment stands and this is shown on the application card.

Experience affects the fit criterion, so the software determines it from a strict order of sources — the first one found wins:

1. total experience stated explicitly at the top of the CV (the first 2000 characters), such as “Total experience: 12 years”;
2. the value extracted by the model, if plausible (no more than 60 years);
3. if the model mistakenly gave months instead of years — converted to years;
4. the sum of the durations of the jobs listed in the CV;
5. if no source gives a plausible value, experience is left blank. A blank is more honest than an invented number.

These are properties of the software, not promises, and each one names the mechanism behind it. Statements 7.1, 7.2 and 7.5 hold by construction and do not depend on how the model behaves.

• 7.1. The model does not know who you are. Before assessment the following is removed: full name, phone number, email address, name of the school or university, graduation year. A name is an indirect signal of gender and nationality, a graduation year gives away age, and a university name signals social background. The model cannot take into account what it never receives. The employer still sees the full CV — only the model's input is de-identified.
• 7.2. One formula for everyone. The software computes the final score and category with the formula in section 4, using the same weights and thresholds for every candidate. The model cannot move a threshold or change a weight for an individual.
• 7.3. Protected characteristics are off limits. The model receives a mandatory rule: judge only job-relevant evidence; do not try to infer, and never use, gender, age, nationality, ethnicity, religion, marital or family status, pregnancy, health, disability, appearance, political views, or place of birth or residence. If the candidate mentions any of these in their own text, the model must ignore it.
• 7.4. Employment gaps and short tenures are not penalised. Project work, parental leave, illness, caring for relatives, relocation, study and military service are named in the model's instructions as legitimate.
• 7.5. Failures do not become low scores. Missing information lowers only the criterion it belongs to. An unusable model answer sends the application to an error state and a new assessment, never to the “Low” category.
• 7.6. Borderline cases are checked twice, by two models from different developers. This reduces the impact of one model's mistake exactly where a mistake would change the category.
• 7.7. Presentation is not scored. Writing style, grammar, CV design and the language the CV is written in are not criteria unless the job explicitly requires that skill. The same information must receive the same score whatever language the CV is in.
• 7.8. A person decides. Under their contract, employers must not decide solely on the basis of the automated assessment. They see the full CV, the answers and the justification for each criterion — not just a number.
• 7.9. Every score can be traced. Each score is stored with the methodology version, the model that produced the final result, whether a re-assessment took place, the first-pass score and the total the model itself suggested.
• 7.10. Candidate data is not used to train AI models.

An honest account of the limits is part of the method, not small print:

• AI models make mistakes: they can misread a CV, misunderstand context or overrate keyword matches.
• Models are trained on text that contains social stereotypes. De-identification and the rules in section 7 reduce their influence but cannot remove it entirely — for example through company names or style of writing.
• Results are not fully reproducible: assessing the same CV again can differ by a few points.
• No independent statistical study of bias on real applications has been carried out yet. It will be done once there is enough data, and its findings will be published here.
• Only what is written is assessed: a skill not mentioned in the CV or answers is not counted. The “Measurable results” criterion favours CVs that express results in numbers.

• Upload a PDF with real text (exported from an editor or a job site), not a scan or a photo.
• Name your skills the way the vacancy names them, if you really have them.
• Give results in numbers where you have them: deadlines, volumes, growth.
• Answer the vacancy's questions specifically, with examples from your own experience.
• If total experience matters for the job, state it in one line at the top of your CV.

You can ask for your application to be reviewed by a person without the automated score, ask for an explanation of the score by criterion, object to automated processing, and ask for your data to be deleted. Contact the employer or write to support@hrmate.app — we pass the request to the employer and act on it on our side.

Any change to the criteria, weights, thresholds or rules is published on this page with a new version date. Every score keeps the version it was calculated under, so scores from different versions are never silently mixed.