Research
What we found, and how we found it
Long-form research with the method attached. Every headline figure carries a receipt saying whether it was measured, computed or cited — including the ones that turn out to be ours rather than the paper's.
Standing measurements
The pages that get re-measured
- The Judgment Index Edition 1, 2026 Q3. What employers are explicitly asking for, read two ways from 2,151,213 active job postings, with both term lists and the k-anonymity floor published. Nothing on it describes a person.
- Benchmarks The AI-usage requirement in live postings, measured by two instruments that disagree by roughly six-fold. Both ends are published and neither is presented as settled.
- The evidence ledger What has to be true before a score is released, which laws bind an automated employment decision in each jurisdiction, and the source behind every figure quoted anywhere on this site.
- Published benchmarks The corpus benchmarks: every published cell with the instrument that read it, the limits it carries, what was suppressed, and an ungated CSV beside each page.
How to read these
Every number shows its working
Each post ends in a receipts appendix: every figure quoted in the prose, and whether it was measured by us, computed by us from figures someone else measured, or cited from a source we have read.
That distinction is not pedantry. The first post exists partly because its most quotable statistic — an 84% higher error rate — was a relative difference we derived and printed directly beneath somebody else's citation, where it read as theirs.
A figure without a receipt does not ship
The build fails a post whose prose quotes a percentage the appendix does not carry. It is a gate rather than a habit, because a habit is what fails on the busy week.
Subscribe
New research, when there is any
No newsletter cadence, no drip sequence. An RSS feed that updates when something is actually finished.