Recording a provider term
Entering, by hand, what a provider published about how it handles the work: retention, training use, processing regions, whether a key you hold is accepted. Each reading is taken from a page on the provider's own domain.
npm run term:add -- --provider openai --kind zero_retention --value yes --on 2026-08-27 --source-url https://openai.com/enterprise-privacy| Flag | Meaning |
|---|---|
--provider | The provider it is about |
--kind | One of zero_retention, no_training, regions, byok |
--value | yes or no, or comma-separated region codes for regions |
--on | The date somebody read the page. Never a date in the future |
--source-url | The page it was read from, on that provider's own domain |
--note | What the page scoped it to, if it scoped it |
The four refusals it inherits
They are the curated log's, applied through the same function:
- No source URL, no row.
httpis not a source.- A host that is not the provider's is not the provider.
- A URL that does not answer 200 today is not verified today.
Plus two of its own. A value that is not one the kind accepts is refused rather than stored, because "the term answers whether the provider trains on customer data, and a hedge is not an answer this record can hold". And a second reading of the same kind on the same day is refused, because a term that changed is a new row on the day it was read, never an edit of the one a decision was taken under.
Why a person and not an adapter
A policy page is prose and the answer is a yes or a no. A parser deciding what "we do not retain your data" means would be interpolating exactly where the honesty contract forbids it.
What it is for
A workload requires one, and a candidate whose provider does not meet it is excluded before anything is weighed:
npm run aperture:workload:edit -- acme/support --constraint requiresZeroRetention=trueProvider term is exact about what a requirement means when the record holds no term to check it against, which is the case worth reading before you rely on one.
