Defensibility
We score the requirement. We never score the person.
Your client’s counsel will ask how the ranking works, and your answer has to survive the question. This page is that answer: what the tool does, what it refuses to do, and — the part most vendors leave out — what it does not discharge for you.
Start here
We had this feature. We deleted it.
Before any of the below, the thing worth knowing about how we think.
An earlier version of this platform scored every person against every role and drew the result as a ring on their card. The portal sorted human beings by it. It looked exactly like what buyers ask for.
When we took it apart, roughly four fifths of that number was a language model’s opinion, blended with an employer weight vector and three defaults that had simply been chosen — 0.4, 0.7 and 0.5 — and then presented to operators as a measurement. We deleted it, and we wrote a build gate that fails the deploy if anything of that shape ever comes back.
When the request came again — score candidates against a job description, make it a flagship — we did not refuse it and we did not quietly rebuild the deleted thing. We built the version that survives the objection, and the four properties below are what make it a different thing rather than the same thing renamed.
How the ranking works
Four properties. Drop any one and it is the tool we deleted.
These are not settings and there is no plan that turns them off.
Every point comes from a requirement you typed
Not an inferred trait. Not a personality read. Not a proxy for anything. You paste the req, we show you the requirements we read out of it in your own words, and you edit that list before anything is scored. A candidate earns points only where a requirement you kept meets a fact they stated about themselves.
The rubric is yours, on screen, and editable
Change any weight, mark anything required, delete a line we should not have picked up. This one is not a nicety: if you cannot change the rubric, then it is OUR rubric being applied to YOUR hiring decision, and that is a different product with a different legal character.
Unknown is not a gap, and they never render the same
If we were not told whether someone holds a certification, that scores zero — and it says we were not told, in different words and a different colour from someone who told us they do not hold it. Those two facts lead to different next actions, and collapsing them is precisely how a ranking tool quietly becomes a rejection tool.
It ranks. It never rejects
There is no auto-advance and no auto-reject anywhere in it. Missing-required is shown to you, and nothing is ever done about it on your behalf. That is the line between decision support and an automated decision, and it is the line the law itself draws.
What it refuses to do
The list of things we will not build, for anyone.
Each of these is enforced by a check that fails a deployment, not by a policy document that fails a person.
No score on a person
No match percentage, no readiness number, no ordinal, no ring. A build gate fails the deploy across the entire codebase if one appears, on any surface, for any audience.
No language model in the arithmetic
An AI helps read your job description into a list of requirements, which you then correct. It touches nothing downstream. Every number is deterministic and rebuilds from its line items or reports itself broken.
No narrowing a search by geography around a person
A radius can be widened and never narrowed, because a geographic filter pointed at a person is a proxy. Every widening is recorded.
No minors, at the read
Minors are excluded from ranking entirely — not reduced into it with fewer fields. They are absent from the query, guardian-gated, never contactable and never in an export.
No fabricated field on a real person
An empty value stays empty. We never fill a blank with a default, an average, or a guess, and no meter is ever drawn from a number we did not measure.
What this does not do for you
Using us does not make your firm compliant, and we will not tell you otherwise.
A transparent, editable rubric is still capable of substantially assisting a screening decision. Where a law governs automated employment decision tools — New York City’s Local Law 144 is the one most firms meet first — obligations such as an independent bias audit, published results and advance notice to candidates are operational duties that sit with the employer or the firm making the decision. They are not something a vendor discharges by writing better software, and we are not going to imply that we have.
The same is true of validity evidence for a selection procedure under the federal uniform guidelines, and of the professional testing standards that apply to any score used to make decisions about people. Our design choice is to keep the tool on the decision-support side of that line and to give you a record you can actually hand over. What your firm owes in your jurisdictions, for your clients, is yours.
Take this page to your own counsel. It is written to be read by them, not around them — and nothing on it is legal advice.
What you can actually produce
The record, when someone asks for it.
Defensibility is not a claim on a website. It is whether you can show your work a year later.
The rubric, as you ran it
The exact requirement list used for that search, with the weights and required flags as they stood — not as they were later edited.
The arithmetic, per person
Every score rebuilds from its line items, including the misses and the unknowns, each one traceable to a requirement you typed.
The submission ledger
Who was submitted to which client, on what date, and what happened next — timestamped, and not editable after the fact.
None of this is generated after the fact or reconstructed on request. It is what the desk writes down as you work, which is the only version of a record that is worth anything when it matters.
Nothing on this page is legal advice