Model card · Resume Assessment · Updated October 2026
ZeroFit
ZeroFit does the first résumé review for you. It reads every application against the role, the same way every time, and ranks candidates by fit so your team spends its time on the people most likely to be hired. This page shows how it works, how well it does, and the safeguards you keep.
3 in 4
Puts the right person first
Pick one candidate a hiring team moved forward and one it didn't. ZeroFit scores the one moved forward higher 74% of the time.
+29%
Better than keyword matching
Ranks the right candidates higher than a standard keyword search does, on the same roles and résumés.
47%
Of hires in the top 10
Nearly half the people who were eventually hired sat in ZeroFit's top 10 for their role.
Measured on roles and candidates almost entirely absent from training. Details below.
How it works
ZeroFit is a language model trained on past hiring decisions. For each role, it looks at the résumé text the candidate uploaded and does three things.
Step 1
Personal details come out
Before scoring, we remove the name, pronouns, contact details, address and labelled fields like date of birth, gender and ethnicity.
Step 2
Copied text comes out
Anything copied word for word from the role description is removed, so pasting in the job ad earns nothing.
Step 3
The résumé is compared with the role
ZeroFit matches on meaning, so experience described in different words still counts.
Result
0–100
A fit score. Higher means a closer match to this role. It isn't a probability of being hired.
How well it works
We test ZeroFit against what hiring teams actually decided, using roles and candidates almost entirely absent from training. Every new version has to beat the current one on this test before it ships.
How high the right candidates land in each role's ranking. Longer is better.
Out of every 10 people eventually hired, about this many were in ZeroFit's top 10.
47% of hires were in the top 10 for their role.
You stay in control
ZeroFit orders the queue. Your team keeps every decision that matters.
It never rejects anyone
A person makes every rejection. The platform refuses any rejection based on an AI assessment that a person hasn't authorized.
You set the bar
You choose the score at which candidates can be invited to interview, and whether invitations go out automatically.
Every candidate gets the same review
The same résumé and role always get the same score. No résumé is skimmed late on a Friday or left at the bottom of the pile.
Every score is on record
Each score is stored with the model version and the threshold in effect at that moment, so you can always see what a candidate was judged against.
Updates are checked before they ship
After every update, known résumés must keep their exact scores. If anything shifts, the release stops.
It doesn't guess
A résumé or role description too short to judge is marked as not scored. It isn't given a low score.
Fairness
We take a résumé, change one thing that shouldn't matter, and check the score stays the same. ZeroFit doesn't see these details at all, and in our tests changing them left the score unchanged:
- Names associated with different ethnicities and genders
- Pronouns
- Age and year of birth
- A career gap for caring
- Spelling slips and non-native phrasing
- Where the candidate lives
These checks use a small set of test résumés, so they rule out large effects rather than small ones. One test with a disability mention moved a score by about one point out of 100. Full results are in the technical details below.
Good to know
- A clear role description helps. ZeroFit matches against the description as written, so a specific one gives a better ranking.
- It reads, it doesn't verify. It can't check whether a résumé's claims are true or see anything left off it. The interview stages are for that.
- Compare scores within a role. A score ranks people for the role they applied to, not across roles.
Questions and human review
Candidates can ask a person to review any automated assessment, and for an explanation of it, through Your data & privacy or at privacy@zerohiring.com. How candidate data is used to improve our models is set out in section 4 of our Privacy Policy. Employers with questions about ZeroFit can use the same address.
Technical detailsShowHide
- Version
- ZeroFit 1. Trained April 2026, in production since May 2026.
- Architecture
- Bidirectional transformer text encoder, fine-tuned to match résumés to roles.
- Trained on
- Past hiring outcomes: which candidates were hired for each role.
- Reads
- Résumé text with personal details and copied role-description text removed, and the role description. Very long documents are truncated.
- Output
- Similarity between the two texts, scaled to 0–100 and stored with the model version and the threshold in effect.
- Speed
- 1.3 s typical and 2.9 s at the 95th percentile per scoring request on live traffic (2,667 requests, 14 days to 30 Sep 2026, none failed). One request scores a batch of résumés for one role.
Accuracy
Held-out test set of 1,820 applications from roles and candidates absent from training, measured August 2026. "Moved forward" counts candidates hired, approved by an interviewer, or progressed past screening and turned down for reasons other than qualifications (94). About 7% of test applications also appeared in training data, which may favour ZeroFit slightly.
| Measure | Labels | ZeroFit 1 | Keyword (BM25) |
|---|---|---|---|
| ROC-AUC | Moved forward (94) | 0.74 | — |
| ROC-AUC | Hired (47) | 0.77 | — |
| Mean average precision | Moved forward (94) | 0.229 | 0.178 |
| Recall at 10 | Hired (47) | 0.47 | — |
Measured before personal-detail removal (30 Sep 2026). We'll republish these after re-measuring.
Fairness
104 counterfactual pairs built from test résumés, each changing one attribute, measured 1 October 2026 with personal-detail removal in place. 102 held: names 28 of 28, pronouns 12 of 12, and every age, caregiving-gap and writing-style pair. One disability mention moved a score by 1.1 points, and one gaming test still lifted a weak résumé; we're tracking it. Location was checked separately on live scores in August 2026: an average of 54.1 for candidates in the New York metro area against 54.0 elsewhere. We don't collect candidates' demographic data, so we haven't yet measured selection rates by group on real outcomes.
Reliability
The same résumé and role description always get the same score. After every update, a known strong résumé and a known unrelated one are scored against a fixed role (78.2 and 36.0); if either moves by more than a point, the release stops.
Version history
- May 2026. ZeroFit 1 goes into production.
- August 2026. Three retrained candidates were tested on the held-out set. None beat ZeroFit 1, so none shipped.
- 30 September 2026. Personal details are removed before scoring. Scores from before and after shouldn't be compared within one role.
- Next. A head-to-head with general-purpose AI models scoring the same résumés, and a re-measure on live hiring outcomes.