Fairness
We test every update for fairness. Here's how, and what the simulations show.
Blind cards stop bias at the moment of decision. But any ranking can carry bias in what it rewards, so we test the ranking itself: with the 4/5ths rule on every update, and with simulations that compare it with the ways people usually hire and network.
Three layers
Blind decisions
Nobody sees a name, photo, age or gender before saying yes. Degrees are never scored. The model never reads a protected trait.
Every update tested
Changes to how the matching counts things are tested by group with the 4/5ths rule before going live, and blocked if they'd push any group below the line.
Never a gate
The app ranks and suggests. It never rejects anyone automatically. People make every decision.
Simulation, not evidence. These are simulation results on synthetic people, not evidence from real hires. The simulator runs the app's real matching code, but the people, and how well each signal predicts real work, are assumptions borrowed from published research. Only a real pilot can say whether it works.
Hiring simulation: who gets hired, and how good the hire is
2,000 simulated hiring rounds, 30 acceptable applicants per job. Bad hires are the bottom 20% of true performance; great hires the top 20%. Group ratio: the hire rate of a group with less education and fewer years on paper, at the same true skill, divided by everyone else's. ⚠ = below 0.80.
| Item | Bad hires | Great hires | Still there at 1 year | Group ratio |
|---|---|---|---|---|
| Random pick among acceptable applicants | 19.8% | 20.4% | 76.3% | 0.99 |
| Résumé ranking (years, education, keywords) | 12.0% | 29.9% | 77.8% | 0.22 ⚠ |
| Résumé shortlist + unstructured interview | 10.5% | 32.9% | 78.5% | 0.32 ⚠ |
| Résumé shortlist + structured interview | 6.2% | 42.6% | 80.0% | 0.35 ⚠ |
| Referral from the manager's network | 18.1% | 22.2% | 80.9% | 0.54 ⚠ |
| Zydeby Work | 6.7% | 40.5% | 83.0% | 0.91 |
Source: the hiring simulator's report, generated Nov 5, 2026. The structured-interview row is a well-run best practice; in the same simulation it did slightly better than the app on performance, and the app did better on retention and fairness.
Fairness on three splits (hiring simulation)
Each split compares two simulated groups with the same true skill. 1.00 means equal hire rates; below 0.80 fails the 4/5ths screen.
| Item | Ratio | 95% range |
|---|---|---|
| Less education and fewer years on paper, vs. everyone else | 0.91 | 0.88–0.94 |
| Age 40 and over, vs. under 40 | 1.06 | 1.02–1.10 |
| Smallest professional networks, vs. the rest | 0.87 | 0.83–0.90 |
An earlier version of the matching failed the network split: people with small networks were hired at about half the rate (0.50), because vouch counts helped people who know more people. We changed how vouches are read so a count never raises anyone's rank, and the ratio came back to 0.87.
Networking simulation: access to good contacts
80 simulated runs of one metro of 120 members over 12 weeks. Access = members with at least one good new contact by day 84. No ratio fell below 0.80, for mentoring, peers or trade partners.
| Item | Ratio |
|---|---|
| Smaller group vs. larger group | 1.03 |
| Age 40 and over vs. under 40 | 1.04 |
| Small networks vs. larger networks | 1.06 |
| Newcomers vs. established members | 0.99 |
Source: the networking simulator's report, generated Nov 5, 2026. In the same runs, 7 to 8% of members in each networking lane never met anyone in 84 days, against 35% of mentees in a chat-first, event-style arm.
What the simulations can't tell you
- The people are synthetic. How real people answer, how honest references are, and how much our questions capture are assumptions.
- How well each signal predicts work is borrowed from published research, mostly from larger organizations. It may not transfer to a ten-person contractor.
- Signals are modeled with independent errors. Real ones share biases (a charming candidate does well with interviewers and references), which flatters every method that combines signals, including ours.
- Fairness groups differ only in the ways we built in. A real audit has to use real data, which is why we ask for optional answers and test every update.
Questions
What is the 4/5ths rule?
A rule of thumb used by U.S. regulators: if one group is selected at less than four-fifths (80%) of the rate of another, that's a sign of adverse impact worth looking into.
Do you collect demographic data?
Only if you choose to, in Settings: gender, age (under 40 or 40 and over) and race or ethnicity, each with "Prefer not to say". It's stored apart, used only for app-wide totals, and never read by the matching. Disability is never asked.
Do you audit individual employers?
No. We never run demographic analysis on a single company. Employers are held to follow-through instead: backing out after the reveal counts against them, whoever the candidate is.
Has this been tested on real hires?
Not yet. The numbers on this page are simulations. A real pilot with outcome data is the only thing that can show whether it works.
Keep exploring
Judged on what you can do.
Free to join. Answer the basics in a few minutes and you can be matched right away. Everything else you add makes matching better.