Meaning / Core concepts

Meaning has context

SUPERB is building a versioned reference for interpreting language. Existing dictionaries provide a starting inventory; contextual evidence can reveal distinctions that inventory does not yet describe.

Passage and target#

A contextual request includes a passage and a target expression within it. The passage constrains the possible readings. The same spelling in two passages need not identify the same sense.

Sense and definition#

A sense is an identity for a distinction in usage. A definition describes that distinction. Editing the description should not silently create a new identity or rewrite the history of an accepted release.

Evidence and interpretation#

A source passage is evidence of usage. A proposed reading is an interpretation of it. A generated explanation does not turn that interpretation into accepted evidence. Keep source attribution and review status visible.

Uncertainty#

An unresolved or mixed reading is a useful result. Deterministic ranking scores are not calibrated probabilities. Applications should preserve alternatives and let a person add context instead of forcing a confident answer.

Typed distinctions#

Meaning, referent, claim, criteria, scope, and strength are different questions. Two statements can use the same word sense while making different claims. A future claim-comparison feature must not relabel word-sense similarity as agreement.

Semantic releases#

A release records a particular semantic reference with provenance and review information. Clients should preserve the release identifiers returned by the service. Do not fabricate a release identity when a response only refers to the bootstrap reference.

Sense prediction and consensus#

predictSense() ranks a sense in context. predictConsensus() reads chain-grounded protocol state. A model output cannot finalize consensus, and a finalized protocol record does not establish universal linguistic truth. Replay verification is a separate check; see the protocol specification.

Private material#

Reading, drafts, audience comments, and organizational documents do not become public contribution data merely because an application analyzes them. Collection, analysis, display, redistribution, and training permissions need distinct treatment. See observation permissions.