Building Superb / September 2026
A reference that can change its mind.
How evidence, human distinctions, and versions can help a language reference evolve.
Language changes because people keep using it. A reference that cannot record new distinctions will eventually describe only part of what people are doing with words.
Definitions describe distinctions.
SUPERB begins with existing linguistic inventories. They offer valuable structure and attribution. They are also a starting model. When contextual usages do not fit that model, the useful next step is to gather evidence and ask a better question.
Do these passages use the expression in the same sense? Are the speakers referring to the same thing? Do their claims depend on the same criteria? These questions should stay separate, because a single similarity score cannot answer all of them.
A definition should explain itself.
Consider a definition of crack that says only “the act of cracking something.” It points back to the word the reader wanted explained. Reference now gives that noun a direct description: a fracture or split produced in a solid object. The noun remains separate from the verb meaning to break abruptly under tension, and both keep their original source identifiers.
The same distinction matters for newer language. The everyday sense of cooked can concern a meal; the slang sense concerns being in trouble or facing failure. Reference now carries short, attributed usage excerpts for the slang reading alongside the older inventory. A reported social-media example and a dictionary’s illustrative sentence are labelled differently, because they are different kinds of evidence.
Compare the readings of cooked →
Human disagreement is information.
A split judgment may point to missing context or a distinction that is not stable yet. Forcing every example into agreement would hide that signal. YES, NO, and UNSURE each have a role.
Machines can help find informative comparisons. They cannot turn their own proposals into independent human validation. The record needs to preserve who or what proposed a reading, how it was reviewed, and what evidence supports it.
Change needs a history.
A versioned release lets software refer to a particular state of the reference. Later improvements should add a traceable history rather than quietly rewriting earlier claims. Source withdrawal and permitted use also need to remain visible.
This is the direction behind the semantic platform and LITERATE. In Reference, each meaning has an identifier and an attributed source. Editorial clarifications preserve the original wording, so a clearer explanation does not erase its history.