Explainability
From August 2026 the EU treats AI in hiring as high-risk. Can your screening tool explain a rejection?
We did not design for the regulation. We designed for evidence — and it turns out to be the same thing. This page is what we would want from any vendor, written about ourselves.
Five questions, and our answers
Ask every vendor these, including us. A tool that cannot answer the first one cannot help you with any of the obligations that follow from it.
- 1
Can it list, per candidate, which requirements were met and by what evidence?
Our answer: Yes. Every comparison returns the requirements one by one: met, with the claim that met it and the candidate’s own quoted sentence; or unmet, and still in the list.
- 2
Is the candidate’s original evidence kept unmodified?
Our answer: Yes. The raw source never changes and the claim layer is only ever added to. Our interpretation of a claim is the only thing we rewrite, and rebuilding it cannot touch what the candidate said.
- 3
Does it show what it could not match, or hide it?
Our answer: Shows. A system that silently drops unmet requirements inflates every score; ours keeps them, which is also why our percentages are lower than some.
- 4
Who decides — the tool or a person? And is that recorded?
Our answer: A person. The platform computes and displays; it never rejects, never auto-advances, and never ranks a candidate out of view. Status changes are made by a named user and recorded with the time.
- 5
Can a candidate see, export and delete what it holds?
Our answer: Yes to all three, from their own account, without asking us. Export is one document; deletion removes everything they wrote, leaving only a shape that identifies nobody.
What the system does about each obligation
Transparency toward the person
Candidates know a graph is built from what they write, see every claim extracted from their words, and can correct any of it. Nothing about them is derived that they cannot read.
Human oversight
The output is a comparison with its evidence, not a decision. There is no automatic rejection in the product and no setting that creates one.
Record-keeping
Claims carry the sentence and the date; comparisons carry the requirement list they were computed from; the request log carries who asked and when, under a request mark.
Data governance
One vocabulary, the original wording kept beside it, the time of every claim explicit — and “unknown” stored as unknown rather than guessed.
The one sentence to take away
A percentage you cannot open is a percentage you cannot defend. Everything on this page follows from refusing to produce one.
Questions about the Act and this product
- Is this legal advice?
- No, and we are not your lawyer. This page describes what the system does and records, in enough detail that whoever does give you legal advice can judge it. The obligations themselves come from the Act and from your own role as the deployer.
- Are you claiming to be compliant?
- We claim the system is explainable and auditable by construction, and we show how. Compliance is a property of how a tool is used inside an organization, not a badge a vendor can hand over — which is why the five questions are addressed to us as much as to anyone else.
- What exactly is logged for a decision?
- The listing’s requirements as they stood, the claims that met each one with their quoted sentences and dates, the requirements that went unmet, the resulting percentage, and the status a named person set afterward with its time.
- Does the model make the decision?
- The model reads text and returns a structure. It does not write queries, does not rank candidates, and does not set a status. Python assembles every query from whitelisted names, which is also why a model cannot be talked into asking for something it should not.
- What about candidates who never agreed to be searched?
- They are not searched. A profile is private when the account is created; entering the pool is an explicit, dated consent, and leaving it removes the person from every search immediately.
Judge it on the product, not the page.
Create a company, post one role, and open a comparison. If it cannot show you its working, nothing else here matters.