How this works, including where it stops

The one constraint everything follows from

You cannot compute fit between a person and a company unless both are scored on the same dimensions. Off-the-shelf personality instruments score the person well and say nothing about the company. Employer review sites score the company and know nothing about you.

So the framework had to be built rather than borrowed, and every axis on it had to be one that public evidence about a company can actually speak to. That constraint threw out several dimensions that would have been interesting and unmeasurable.

The eight dimensions

Two kinds of axis. On a congruence axis, distance in either direction costs you: wanting more structure than you get and wanting less are both friction. On a needs-supplies axis, only undersupply counts. A company that protects your evenings more fiercely than you asked for is not a problem, and reporting it as one is how a fit model loses your trust.

• Autonomy to Structure

  Congruence

  How much direction you want.

  Deci and Ryan, Self-Determination Theory (2000). Cameron and Quinn's competing values framework. O'Reilly, Chatman and Caldwell's Organizational Culture Profile.

• Pace & Intensity

  Congruence

  Steady, or high velocity.

  Karasek's Job Demand-Control model (1979). Bakker and Demerouti's Job Demands-Resources model (2007).

• Boundaries

  Needs and supplies

  Where work stops.

  Sonnentag on psychological detachment from work (2012). Job Demands-Resources, for the burnout pathway.

• Directness

  Congruence

  How feedback travels.

  Edmondson on psychological safety (1999).

• Growth Model

  Congruence

  Mentorship, or the deep end.

  Self-Determination Theory's competence need. Kristof-Brown and colleagues on person-job fit as distinct from person-organization fit (2005).

• Collaboration Mode

  Congruence

  Consensus, or ownership.

  Cameron and Quinn's Clan quadrant. O'Reilly and colleagues on team orientation. Schwartz's self-transcendence and self-enhancement axis.

• Stability & Change

  Congruence

  Predictable, or constant reorg.

  Schwartz's openness to change and conservation axis. Cameron and Quinn's flexibility and stability axis.

• Mission & Meaning

  Needs and supplies

  Whether the work has to matter.

  Wrzesniewski and colleagues on the job, career and calling distinction (1997). Schwartz's universalism and benevolence.

Authors and years only, deliberately. Effect sizes from these papers are not printed here because they have not been checked line by line against the originals, and a figure written from memory is a figure nobody can check, including us. An author and a year you can verify in one search.

Measuring the person

28 forced choices between two things you want. No sliders and no rating anything one to five, because almost everybody says they value almost everything. What you give up when you have to choose is the only part that carries information.

The 28 items are a complete round robin: every pair of dimensions is compared exactly once, so each of the 8 appears seven times, four times at one pole and three at the other. That buys equal chances per dimension, no advantage from the order they are drawn in, and a head-to-head item to break every tie.

An axis your answers barely touched is marked as unestablished and asked about directly at the end, rather than reported as a confident middle. A fabricated fifty is worse than an admitted gap.

Measuring the company

Public sources only, in four classes, weighted differently on different questions. A company can credibly tell you what its mission statement says. It cannot credibly tell you whether anybody messages you at nine at night.

• Company-controlled

  Careers pages, job adverts, engineering blogs, press releases. The authority on what a company says about itself, and evidence of nothing else.

• Employee-voiced

  Public reviews and forum posts. The only place some things are visible at all, and the place where one angry account proves the least.

• Third-party

  Reporting and analysis by people with no stake in how it reads. Strongest on things that happen in public, like layoffs and reorganisations.

• Structural

  Filings, published pay scales, policy history, seniority distribution. Slow to change and hard to spin.

Every finding is stored as a claim with a source URL, a verbatim quote and a date. A sentence in a report that cites a claim which does not exist cannot be written to the database, because the citation is a foreign key. That is the guardrail as a piece of structure rather than a promise.

Names are removed at the point evidence is extracted, before anything is stored. No report names an individual, whether an executive or a reviewer.

How evidence becomes a number

The model reads and writes. The arithmetic is code. Scoring is deterministic and unit tested, so identical evidence produces an identical score, and every parameter below lives in one file rather than scattered through the functions that apply it.

Age halves the weight of a claim every 18 months

A review of a company that has since had two rounds of layoffs and a new chief executive is close to worthless. Evidence from before a discontinuity, an acquisition, a mass layoff, a change of leadership, is discounted again rather than averaged in.

Sources that are not independent count once

Eight articles rewriting one press release are one source, and so are five reviews written the same week after the same reorganisation. A cluster of eight contributes about three, on a curve rather than a cliff.

A contradiction needs a gap of 25 points and evidence on both sides

The bars are deliberately different: 1 claim on the stated side, because a careers page is published and deliberate and there is nobody else to ask, and 3 from at least two unrelated origins on the observed side, because one unhappy review is not the workforce's position.

A declared dealbreaker gates the score rather than weighting it

If you say something is absolute, a severe mismatch on it caps the index outright. Two annoyances are not one dealbreaker, and averaging them is how fit models produce numbers people recognise as wrong.

These are design parameters, not research findings. Each is traceable to a stated position about which kind of source can be trusted on which question, turned into arithmetic so the arithmetic can be argued with. None of them is an effect size and none should be read as one.

When there is no number

Below the evidence floor the report gives no fit index at all. It says what was found, says what could not be found, and explains why a score would mislead. An unscored report is a valid output, and for small or private companies it will be the common one.

The floor is cleared only when all of these hold:

• At least 4 genuinely independent origins, not pages.
• At least 4 of the eight dimensions actually covered.
• No more than 80 percent of the weight coming from the company itself. A report built only on marketing is a report about marketing.
• Evidence on at least 2 of your top 3 dimensions. Knowing nothing about what you care most about is exactly where a confident number does the most damage.
• Something found about anything you declared a dealbreaker.

A fit index never renders without a confidence band beside it, anywhere, including in a PDF or a share card. Three of these rules are constraints in the database rather than conventions in the interface, so the interface cannot break them by accident.

The three floors, which never enter the score

Some things are not a matter of fit. They are listed separately, with their sources expanded by default, and the fit index is identical whether or not they are present. Folding them into a number would let a high score bury them.

• Psychological safety

  Whether raising a problem is survivable.

• Discrimination and harassment

  Substantiated patterns, not single allegations.

• Compensation integrity

  Whether people are paid what they were told they would be paid.

Reporting one of these as a finding rather than an open question takes multiple independent, recent, high-credibility sources. It is the highest evidence bar in the product, and most of the time the honest output is the open question.

What this is not

It is not a background check, not a rating of any company, and not advice. There is no company score anywhere in the data model, only person-company pairs, so there is nothing to leak into a league table. Evidence is a snapshot of what was public on the day, and companies change.