ORD-2.2
An AI model provider using it for retrieval augmentation, under ORD 2.2
Trains, fine-tunes, or serves machine learning models commercially. Indexing the Work so a model can retrieve from it at inference time. This resolves to the Canopy tier with a Compute licence required on top.
Tier
Owed
A quarterly Canopy Report. Section 5.1 imposes a royalty only when the Canopy Licensee makes Production Use. Plus a separate Compute Fee before the training run.
Sections in play
1.8, 3.1, 5.1, 5.2, 6.1, 6.2, 6.3, 6.5
Who and what
The situation being resolved.
Actor
An AI model provider
Trains, fine-tunes, or serves machine learning models commercially.
over 2M USD
Situation
Using it for retrieval augmentation
Indexing the Work so a model can retrieve from it at inference time.
Rulings
4 questions resolved against the text.
Which tier applies?
Canopy. Trains, fine-tunes, or serves machine learning models commercially. Section 5.2 reporting applies; the Section 5.1 royalty attaches when the Licensee makes Production Use and only to attributable revenue.
Sections 1.8, 3.1, 5.1
Is anything owed?
A quarterly Canopy Report. Section 5.1 imposes a royalty only when the Canopy Licensee makes Production Use. Plus a separate Compute Fee before the training run.
Sections 5.1, 5.2
Does training need its own licence?
Yes, at every tier including Root. Section 6.1 removes AI Training Use from the base grant entirely. The fee is charged per million records included in the training corpus. Training without it is not a lesser breach of these terms, it sits outside the licence, which Section 6.5 treats as infringement.
Sections 6.1, 6.2, 6.5
What must the resulting model do?
Carry provenance disclosure on output that reproduces the Work substantially word for word, above the threshold set in Exhibit B. Section 6.3 targets the attribution harm alleged against Copilot rather than attempting to police memorisation.
Sections 6.3
Cautions
What this situation gets wrong most often.
Corpus Inclusion under Section 1.11 reaches datasets compiled with actual knowledge they will be used for training. Passing the Work to a third party who trains on it does not move the obligation off you.
Retrieval augmentation falls inside the Section 1.10 definition. Whether that is the right line is one of the published open questions rather than a settled point.
Section 8.3 of ORD requires a withdrawal of consent to propagate to any derivative you have distributed. Whether that is operationally achievable at depth is a published open question.
Same actor
Other situations for this party.
Non-normative. Where this and the licence text disagree, the licence text applies.