A shortlist that survives “why these three?”

UnicornWorks assesses a pairing, not a person. You describe one open role and name the people you are considering. For each pairing it returns how that person would meet that work, what it would cost them to hold it, where the friction shows up, and what would close the distance — then ranks the slate on one shared axis, so your second and fourth choices can be read against each other rather than in isolation.

It runs on the same engine as our Signal Read product, aimed at a different question. Signal Read reads a person in order to reach them. UnicornWorks reads a person against a role in order to decide something about them. That difference is why they are two products with two sets of obligations, and why this one carries a compliance posture the other disclaims.

A résumé tells you what someone has done. An interview tells you how they present. Neither tells you how the work will land in this role, on this team, at this stage of the work — which is the thing that actually decides whether a good hire turns into a good placement.

Where the existing instruments stop

• Drive-based tools match what a person wants, not what they can sustain. The result is a placement that performs well at first and strains when the work changes shape. The strain gets read as a people problem; it is usually a composition problem.
• “Culture fit” as commonly practiced is a vibe with a veto — unfalsifiable, unevidenced, and the most reliable cover story for an unlawful decision. We replace it with cues quoted from your own stated culture, and with no score at all.
• Scored screening tools hide the reasoning and ship the number. The number is the part a hiring panel cannot argue with and cannot check.

So the design commitment runs the other way: the reasoning is never withheld, the footing under every claim is stated, and the product always says how someone could be a fit rather than only how they fall short.

A person is read in six separate categories. Each answers a question no other category answers, each is read on its own evidence, and each declares what footing it stands on.

This is the load-bearing design decision, and it is the opposite of what most tools do. We do not reconcile six readings into one “true” profile. A read that collapses the categories into a single answer has thrown away the finding.

Because the interesting fact is almost never any one category. It is the gap: someone whose published self sits a long way from their delivered self is holding a position that costs them something to hold, and that cost is going to show up in your role whether or not anyone named it in the interview.

Gap classification is arithmetic, not narration: same family reads as aligned, a neighbouring family as flex, a cross-family result as divergent. Computed, not written by a model. Two consequences that matter to a buyer — it cannot contradict itself the way a generated paragraph eventually will, and it costs nothing to run, so it runs on every read including the free one.

A role is not one kind of work. You force-rank at least three of the five phases of work — and may rank all five — because naming only one scores the job as though the rest were irrelevant.

• Initiate
• Coordinate
• Integrate
• Modulate
• Punctuate

How much each ranked phase counts is then derived from your own role description by counting the phase language in it — not set by a model, and not a fixed 60/30/10 applied to every three-phase role. Put the same duties in tomorrow and the same distribution comes out, which means “what produced this ranking?” is answerable by pointing at your own words.

The split, and why both halves stay honest

• Your ranking sets the order. You know the job. A derivation that reordered your ranking because a paragraph came out differently would silently overrule the hiring manager.
• Your prose sets the concentration. How saturated the description is in one kind of work is the part nobody can state as a number — and it is exactly what a fixed template invents.

Bounded three ways, enforced in code rather than hoped for: no single phase can exceed 75% of the weight, no phase you named can fall below 4%, and a lower-ranked phase can never outweigh a higher one. If your description carries too little phase language to derive anything from, we fall back to a stated distribution and the record says that is what happened.

1. Each category is read on its own evidence at the granularity that evidence actually supports — family level where only that is justified, and nothing at all where the material will not carry a claim. A claim requires the inputs it is made of; thin evidence yields no claim rather than a guessed one.
2. The model writes a figure for the pairing, and the path to fit. If the path is missing, the number is withheld. A bare number is the precise artifact this product exists not to ship, and inventing the reasoning underneath one would be worse than showing nothing.
3. A structural ceiling is computed and applied afterwards — the architecture binds last, so a persuasive paragraph cannot write around a structural finding.
4. A margin is derived from the evidence tier and from how well you say you know the person. Certainty caps at 95 and floors at 45 — agreement is never certainty — and the ± band comes from that one source, so the band and the certainty can never contradict each other.

The ceiling table

Where a role names several phases, the ceiling is the weighted blend of what each phase would impose on its own — which is where your ranking actually moves the number. A miss in a role’s pivotal phase is weighted to bite harder than a miss in a supporting one.

We do not withhold the read when evidence is thin. We state how far to trust it. There are three routes to a candidate’s own answers, and the route is recorded with the result.

1. Define the role. Title, what the job actually involves, and the force-ranked phases. Optionally, your culture in your own words. About five minutes, once per role.
2. Add the candidates. The people you are considering, plus whatever signal you have on each — a profile link, a CV, what you know firsthand. A minute a candidate.
3. Send invitations where it matters. Send a role-bound invitation to the candidates you are serious about, or take a code one offers you. This is the step that moves a read from inference to measured. Seconds for you; about fifteen minutes of theirs.
4. Assess the slate. Per-candidate reads, the ranked comparison, and an explicit flag where rows do not rest on the same kind of evidence.
5. Take the probe plan into the interview. Each divergence becomes a specific thing to ask; each thin category becomes a specific piece of evidence to go and get. Generated from the read’s own seams, so the interview spends its time where the record is actually unresolved.
6. Feed back what you learned, and re-run. Debrief notes, work samples, reference conversations. Footing rises, the read sharpens, and the lineage of what changed is kept.
7. Decide, and record what you decided. You decide; the audit row is written either way. Recording the decision beside it takes one click, and it is what turns the log from a record of what a tool said into a record of what a human did with it. Later, how the placement is going goes in the same place.

• Coverage is a union, never an average. It is the set of phases the team’s members naturally cover between them — which is why a five-person team can be worse covered than a three-person one, and why a team mean would have hidden exactly that.
• A phase held from someone’s strongest position is not the same as one held from a secondary. Both are genuinely covered; one is genuinely weaker. The read keeps them apart.
• You get a structural band and the facts that drive it — strong, workable, thin, or not enough measured to say — and no team percentage. A number on a team reads as a grade, and this is a lens.
• An unmeasured roster returns “not enough measured to say”, never a low score. A team nobody has measured is not a weak team.
• None of this can reach a candidate’s score. Team structure is a property of the team, and the slate’s scoring path cannot see it.

Where the read names a pattern — a team routing its stretch work to whoever can absorb it, say — it states that as a tendency of the configuration, not as an observation about what any named person is carrying. What someone actually absorbs is known only to them, and asking is the only way to find out.

You state your culture in your own words. The read is grounded in them: a conservative cue lexicon, each claim carrying the exact phrase it rests on, and an honest “your description doesn’t say” where nothing matches. Under-claiming is the design. A culture read that cannot show you where it came from is the vibe this is meant to replace.

What comes back is standings and paths — what your stated culture rewards, what it taxes, what it does not mention — beside the fit figure.

The method behind the read, in more detail and for both products, is on how a read is made; what we store and for how long is on what we store.

Per open role, so it prices against a search rather than a headcount. At several concurrent searches the workspace is the shape that works.

A buyer brief that only lists strengths is an advertisement. These are the real limits, and we would rather state them than have you find them.

• The wiring category is unavailable until the deeper instrument is in play. It ships declared-unavailable in every read rather than quietly missing.
• The ceilings and the band thresholds are tuning, not architecture. What is structural is the ordering — that a measured mismatch is worse news than no information, that a phase held from a prime counts for more than one held from a secondary. The exact numbers want real roles to tune against.
• Calibration is thin. Most of those numbers were set against constructed slates rather than a corpus of real placements. We will not imply a validation study exists.
• The Culture lens sits at its “stated” rung — your words, quoted back. A measured culture rung is a later build.
• There is no outcome corpus yet. The product records what it said, what you decided, and how the placement is going. What it cannot give you is a hit rate: that needs a body of real placements watched over real time, and the time is not shortcuttable. So nobody, including us, can quote you a hit rate. Anyone who does is selling you something. What we can tell you is exactly what the record holds and the date it started.

We keep a falsification list in the repository and re-derive it at every session close. If any of these becomes true, the product has broken its own spec — and this is the list to hold us to.

1. It returns a verdict — anything a reasonable reader would act on without a human judgment.
2. An unevidenced number stops being labeled, and is presented as though it were measured.
3. The ceiling stops binding, so an unmeasured candidate can score as if they had been measured.
4. A ranking happens with no audit record, or a deletion removes the fact of the ranking rather than just the name.
5. Demographic data is collected from candidates by this product.
6. Consent obtained for one named purpose is used for another.
7. The reasoning is withheld — a score shipped without where it rubs and how the person could fit.
8. It merges onto one surface with our non-hiring product, putting “not a screening tool” on a product where that claim is false.

• A pass/fail screen that thins a pile for you. There is no threshold here, by design.
• A personality test. This measures a pairing, and says nothing about anyone’s worth.
• A culture score. Ruled out, on purpose, and not coming.
• A performance predictor. We do not claim it and would not know how to support the claim. We read a pairing and show the reasoning, so a human can decide better.