A benchmark with a human baseline behind it.
Labs are not short of text. What is scarce is evaluation in a domain where being wrong has consequences, and where you can see what a qualified human actually did on the same item. We have both, and we published the baseline before we offered anything for sale.
Our held-out evaluation set was not private
Correction, 29 September 2026
CRLN-AIE-001 v0.2 described 57 scenarios as a future private evaluation set. Their answers could be read without an account through three routes. We closed two of those routes on 29 September 2026 and have retired the set. The published v0.2 results rest on 114 items that were public by design and are not affected.
One route is still open: the Practice Lab app still sends every option's score to the browser. Until answers are scored on the server, we cannot claim that any held-out result of ours is free of contamination.
v0.3 will use 60 new scenarios, keyed by a practitioner, whose text and answer key never reach a browser, a repository or a public table.
The baseline exists, and it is already public
A benchmark with no human comparison tells you a model scored 3.7 out of 5 and nothing about whether that is good. Ours was measured against clinical research professionals answering the same items, and it is published rather than described.
Models against practitioners on the same item bank, with the method stated. Version 0.2 publicly reverses the conclusion of version 0.1.
Where the assessed workforce actually is, by country and role.
What is licensable
Availability is stated next to each one. An asset that is not ready is a roadmap item and is labeled as such.
- The evaluation item bank: rubric-scored clinical research decisions with citations to the governing clause, usable to test whether a model makes a regulated judgment call correctly. Available once items discriminate. Rewrite in progress.
- A held-out set of 60 new scenarios whose text and answer key are never sent to a browser, never stored in a repository, and never placed in a public table. It replaces the earlier held-out set, retired on 29 September 2026 (see the correction above). Not yet written. Not yet offered.
- A custom evaluation harness built against your own task definition, using our domain and rubric rather than ours off the shelf. Available now.
What is never licensed
Stated here rather than in a contract annexe because your procurement team will ask, and the answer is better given before it is asked.
- Row-level learner records, at any price, under any agreement.
- Learners' free-text reasoning in their own words.
- Anything from which an individual learner could be identified, including small aggregates. Nothing below five people leaves the platform.
Why this is difficult to replicate
This data exists because people practiced and were graded against a published rubric, not because anything was scraped. A competitor with substantial funding cannot buy it; they would have to run a training platform for years, give it away free, and earn the participation first. That is the whole asset, and it is why the baseline is published: an evaluation nobody can check is not an evaluation.