AI Optimization Terminology
What standard means here, and what it does not
The word appears constantly in this field and is doing at least three different jobs, only one of which involves anybody agreeing to anything.
Three things called a standard
A published specification. A technical document defining how something works, produced by a body with a process. Implementable, testable, and it exists whether or not anyone adopts it.
A widely adopted convention. Something enough parties do that it becomes the practical default, with no document and no process. Real, powerful, and it can change without notice.
A recommended practice. Something a party asserts is the right way to do something. This is what most uses of standard in AI Optimization Terminology actually mean, and it involves no agreement by anyone else.
The question that separates them: who agreed to this, and where is that recorded? A specification has a document. A convention has measurable adoption. A recommended practice has an author with an opinion.
Why the conflation is convenient
Calling a recommended practice a standard borrows authority that was never granted. It is rarely a deliberate lie. Someone genuinely believes their practice is correct, standard is a natural word for correct, and the implication of consensus arrives free.
The effect on a buyer is real: standard implies risk in not doing it, and non-standard implies deviance. Both are unearned when nobody agreed to anything.
How we use the word
A published specification is named with its publisher and version. An adopted convention is described as a convention with whatever adoption evidence exists. A recommended practice is called a recommended practice with the recommender named, including when the recommender is us.
Where adoption is unknown, that is stated. An unknown adoption level is an honest and common finding for anything in this field published in the last two years.
The related word to watch
Protocol. It has a precise technical meaning and is increasingly used loosely for any repeatable process. A protocol that no second party implements is a method. Both are legitimate; only one implies interoperability, and interoperability is the whole reason anyone cares.
Why this matters more here than elsewhere
This field is young enough that almost nothing has gone through a standards process, so nearly every use of the word is the third kind. A reference that repeated them all as standards would be laundering opinions into consensus at scale.