Release 2026-08-01 — AI-for-Science baseline
31 organizations, 60 sources. The first build. Certified DO NOT PUBLISH, published anyway, with the verdict printed beside it.
Release notes for the original 2026-08-01 build of what was then the AI-for-Science Ecosystem Map. Superseded by r2 the same day. Kept live because r2 carries its entities forward.
Superseded by 2026-08-01-r2 on the day it shipped. It stays live because r2 carries its entities forward, and the reasoning that produced them should remain checkable.
Open the map ↗ · Certification report
What it was
31 organizations placed across ten workflow stages, from literature grounding through clinical and regulatory translation. 27 dated developments. 60 linked sources, each tagged with what kind of evidence it is.
The differentiator was never coverage. It was that the source class travels with the claim, so a reader can see which entries rest on something a third party checked and which rest on a homepage.
Certification
Verdict: DO NOT PUBLISH — 2 fail, 8 warn. Published anyway, with the report beside it.
Failures:
- [E1] 32 of 60 sources (53%) carry no date.
- [E6] 2 entries dated on the snapshot date were the map re-tagging its own cards, counted alongside events in the world.
The warnings are more interesting than the failures, because they are about the taxonomy rather than the data:
- [T2] Mean 4.81 stages per organization out of 10 — the average entry was claimed to occupy a third of the lifecycle, which means the taxonomy was not discriminating.
- [T4] One stage held 23 of 31 organizations (74%). A stage most entries qualify for measures label elasticity, not market structure.
- [T3] Four organizations were tagged into more than 60% of all stages, usually a placeholder for large incumbent, not researched.
- [S3] No stable IDs, so month-over-month diffs fall back to name matching, which cannot distinguish a rename from a removal plus an addition.
- [E4] 25 of 31 organizations (81%) had no corroborating source beyond their own domain.
- [E7] 19 of 27 developments (70%) fell in a single month — a recency-weighted search pass, not a trend.
- [G1] 7 gap claims over territory the map never surveyed.
Those warnings are why the stage inclusion tests exist, and why no count from either release is quotable until they are applied.
What r2 fixed
Stable entity and source IDs (S3), separated classification changes from events (E6), and added explicit verification status. It did not fix the undated-source problem — that ratio got worse, from 53% to 87%.
Comments