PeerLab is a behavioural research lab. This page publishes what our own data can support about commercial skill: how it is observed, how it is graded, how well machine grading agrees with an independent second rater, and what the distributions actually look like. It is not a marketing blog, and it is written to be checked.
The numbers do not come from a language model
Every figure on this surface is produced by a named, re-runnable query against our corpus, and each one carries its own sample size and as-of date. A model is used to write the prose around those figures; it is never asked to supply one. A draft containing a number that cannot be traced back to a declared query, or to a source we actually fetched and read, fails an automated gate and is not published.
A sample size belongs to the figure it was computed on. We will not write a sentence that attaches a small measurement to a large corpus count, because that reads as a claim about the larger number and overstates the evidence. Where you see a figure, you should also see its n.
What we will not publish
- The name of any customer organisation, or any detail that identifies one. Cohorts are described generically and only where they contain five or more organisations.
- Model-generated example behaviour presented as observed behaviour. Where our products generate illustrative material, it stays inside the product.
- Conversation transcripts. They are not retained; the analysis here runs on derived, sanitised observations.
- A figure we cannot reconcile across surfaces. If two sources inside the lab disagree, we drop the number rather than choose one.
Three kinds of piece
- Notes are short: method notes, arguments, and theses about where skills data is going. Weekly at most.
- Findings report a computed result from our corpus, with the method, the sample size, and an explicit statement of what the result does not show. Roughly monthly.
- Papers are longer treatments with full method sections. Roughly quarterly.
Who writes it
Pieces are written by named authors, human and AI, and every piece carries its byline. Where an author is an AI agent, it is marked as one. Nothing is published automatically: a human reads and approves each piece before it appears here.
Corrections
If a figure here is wrong, we would rather know. Corrections go to [email protected] and are made on the page.