The corpus
Every instrument shares the same definitions, Root and Canopy structure, separate Compute licence, and permanent-term continuity rule. The common spine reduces repeated review, but each artifact-specific grant, carve-out, Section 8, and exhibit still needs to be assessed.
Source code
Strip-mining, relicensing reversals, and untracked AI training.
Redis and Valkey, 2024 to 2025. Redis left open source in 2024, was forked as Valkey, and reversed a year later after Valkey had already reached 83 percent enterprise adoption.
Datasets and corpora
Most established dataset and content licences do not state a complete training-consent and compensation mechanism.
Published licence text, 2025. Current standards work on AI training preferences shows that existing licence and metadata practices leave important consent and signaling questions unresolved.
Model weights
The open-weight trap. User thresholds, regional exclusions, and output-training bans, all presented as open source.
Published licence texts, 2025. Widely distributed weights ship under terms carrying user thresholds, regional exclusions, and field-of-use limits, while being described to the public as open.
Skills, prompts, workflows
The widest gap of the four. Packaged agent expertise ships under plain MIT or nothing at all, because no licence was ever written for it.
Surveyed August 2026. n8n workflow templates ship under plain MIT, and no licence anywhere distinguishes packaged agent expertise from ordinary source code.
no fee, no reporting
Individuals, nonprofits, schools, governments, and companies under the threshold
of revenue above the threshold, capped at 250,000 USD per year
Companies above the revenue threshold the licensor set, twenty million United States dollars by default
paid before the training run
Anyone training an AI model on the work, at any tier
A licence is written against named, sourced failures rather than principles. Every clause has to point at something that already went wrong in public.
The draft is attacked from four positions before publication: adopter, commercial licensee, maintainer, and hostile sceptic. This is structured review, not external legal counsel.
The text is published with a digest and its unanswered legal questions. Publication freezes the bytes; it does not imply counsel approval or remove the disclosed legal risk.
Comments, objections, and steward responses are recorded against the published version. An accepted correction can only appear in a new numbered version.
External legal review remains outstanding. If counsel identifies a defect, the published text stays frozen and the correction ships as a later version.