We scanned 40 public semantic models from GitHub with the Semanticus AI-readiness analyzer: Microsoft's own samples, well-known community tools, real production repos and training material. Scans are offline and read metadata only. Every repo is pinned by commit, and the raw per-model scores are downloadable below, so anyone can re-run the scan and check us.
Every official Microsoft sample in the corpus graded F. These are not broken models; they render their reports fine. They fail the things an AI consumer needs: descriptions, unambiguous names, synonyms, linguistic schema, Prep-for-AI settings, documented relationships.
Disclosed bias: public, source-controlled models are the polished end of the ecosystem. Authors published them for an audience, and teams using git are the engineering-mature minority. Typical private client models are rougher than this sample, so the median here is plausibly an upper bound.
Download the raw scores (JSON) · scanned
4 July 2026 · scanner and corpus are in the Semanticus repository under tools/readiness-corpus.
The claim behind Semanticus Pro is that an AI assistant wrapped in the engine's verify/probe/benchmark loop writes more-correct DAX than the same assistant retrying on raw error text. We are measuring exactly that, and we pre-registered the method before running it:
Results, raw run logs and full session transcripts will be published on this page.
Semanticus