Methodology
One standard behind every PYX Labs benchmark — built by the I-O psychologists who do this work every day.
"Employees don't always speak up, and when they do, whether it leads to anything depends on how well the organization hears them. That makes employee feedback some of the highest-stakes, most easily misread data there is — tied to people's livelihoods and to who holds power at work. We need to ask not whether a model sounds capable — we need measures that assess whether it meets the bar that expert practitioners actually hold and what executives need to make better decisions."
Every PYX Labs benchmark tests whether frontier models can do the real work of a workplace practitioner: understand the human experience of work and provide science-backed recommendations.
Five steps, from real data to a defensible score.
Tasks from the real job
Built from a formal I-O job analysis and a panel of practicing experts — what the job actually demands, not ad hoc prompts.
Real data, real organizations
De-identified datasets from real organizations with real people.
Expert gold standard & criteria
An expert completes each task and defines the criteria for a good answer.
Calibrate the judge
Experts grade sample outputs to tune the AI judge. We know what "good" looks like, and our judge does too.
Identical model runs & scoring
Every model gets the same brief and the same data. A different-vendor judge scores each answer against the criteria.
PYX Labs Benchmarks align to three pillars of performance.
Accuracy & data integrity
Are the numbers right. Does it avoid hallucinating. Does it catch anomalies in the data.
Technical & domain expertise
Is the analysis grounded in behavioral science. Does it handle sensitive comments responsibly. Is it contextualized to the industry.
Communication of insights
Is it coherent, executive-ready, specific, and tied to stakeholder priorities.