Our values for
better recruiting.
Four things we stand for. Not a poster in the office — each one is a mechanism you can open, test and challenge inside the product.
Where we could measure them, we did — and published the result, including what it does not prove. Read the CV study →
Blind where
it counts.
We do not claim our models are free of bias — nobody honestly can. We take away the fuel instead. What the scoring never sees, it can never discriminate on.
We measured what that filter actually removes: 97.6 percent of the detectable gender signal in our own test set. The remainder is small — not proven to be zero, and we say so. How we measured it →
Two passes, both before the first score. A deterministic rule pass removes names, e‑mail, phone, age and salutations. A second, separate model pass removes what rules cannot catch: origin, gender, family status, health, appearance. The raw transcript is kept untouched for the human record — only the redacted version reaches the evaluation.
“Hi, I'm Lena Berger, 34, originally from Poland — I rebuilt the payroll pipeline at my last employer and cut the error rate by 40 percent.”
Name, e‑mail, phone, age, birth year, gendered salutations. Deterministic, testable, always on.
Origin, gender, religion, family status, health, appearance. Only affected words change — nothing is rewritten.
“Hi, I'm removed, removed, removed — I rebuilt the payroll pipeline at my last employer and cut the error rate by 40 percent.”
Process ownershipStrong+
“I rebuilt the payroll pipeline and cut the error rate by 40 percent.”
CommunicationGood+
“I walked our works council through the new process before rollout, and we adjusted it together.”
Tooling depthOpen+
Not enough evidence in this conversation. Flagged as an open question for the next round, not guessed.
Every output
is traceable.
A number without a reason is a verdict. In Sprad every rating unfolds into the evidence behind it: the sentence, the source, the moment it was said. No evidence, no score.
A quote only counts if it is found literally in what the person actually said — anything the model cannot back up is discarded before it can lift a score. Each criterion is scored three independent times and the median is kept, and the total is added up in code, never by the model. That makes a rating reviewable by your team, defensible in front of a DPO or works council, and contestable by the person it is about.
AI advises. Humans decide.
The AI reads, structures and drafts. People make every call that touches someone's livelihood — at any volume, with their name on it. That line is architecture, not policy.
Invite to interviewconfirmed by Anna
Offersigned by JonasRejection, with feedbackreviewed by AnnaA rejection is the one effect you cannot take back, so there are exactly two ways to trigger one, and both carry the name of the person who decided. A guard test in our codebase scans every source file and fails the build the moment a third path appears — nobody can quietly add an automatic rejection later.
People own
their data.
A profile belongs to the person in it. They see what it holds, release it item by item, and can take the release back — without asking anyone.
Every release is scoped to one company and versioned, so a person can always see who holds what. Withdrawing it revokes the release across the whole company at once and is written to the record. Nothing about a person is bought, scraped or merged behind their back, and their data is never used to train AI models — that is in the contract, not just on this page.
The floor
we stand on.
Values only count when they cost something. This is the concrete under ours.
What we will never build.
Honesty over badges. We would rather tell you exactly what a certificate covers and what a model cannot do than decorate a footer. If a claim on this page ever stops being true, the feature goes — not the value.
Audit us live.
The fastest way to check these four values: watch the AI work. Traceable scores, release screens and human approval gates, all in the product.