A values-illustration definition of art, scored against a verbatim-grounded corpus

Abstract

This is a short record, not a full paper. It states a definition of art, gives its provenance, and reports how well it fits a corpus of cases taken verbatim from the literature. The definition. “An object, action or product created or chosen deliberately to illustrate values.” Two earlier formulations by the same author are also scored. The three differ in two places and both differences matter: moving from object to object, action or product admits performance, ritual and process-based work, and moving from created to created or chosen admits the readymade — the case on which most necessary-condition definitions of art break. Provenance. The core of the definition was formulated by the author as a secondary-school student, years before the project reported here and independently of the literature it is scored against. A one-sentence definition is not the sort of thing copyright protects; what can be established is priority, a dated public record bearing a name, which is what this deposit is for. Method. Rival definitions of art are treated as classifiers and scored against cases taken from what papers assert, each anchored to an exact quotation from the paper rather than a paraphrase or a survey response. Cases are labelled three ways — asserted to be art, asserted not to be, or raised and left undecided — and undecided cases are excluded from the metric rather than counted as either answer. The corpus is 52 tagged papers, of which 29 carry verdicts, giving 660 cases of which 355 are adjudicable. Agreement is reported as Matthews correlation rather than F1, because about 80% of cases are positive and F1 ignores true negatives. Intervals come from a paper-clustered bootstrap. Result. The definition scores MCC +0.662 (95% interval +0.525 to +0.801), placing it among the leading definitions and above every classical account except the cluster and open-concept theories. Its interval overlaps those above it, so the claim is “among the leaders” and not “the best”. A deliberately circular control — “whatever people call art” — scores +0.983, which is the calibration check: had it scored materially lower, the coding would contradict itself and no other number could be read. What this does not establish. The cases were coded by a single coder, so two-coder reliability is not met and is not claimed. The measurement is corpus-relative by design: change the corpus and the ranking changes, which is the finding rather than a caveat. Scores are comparable only within one run, since instruction wording alone was measured to move MCC by about 0.12 for an identical definition. The corpus tilts contemporary. Reproduction. Every number is recomputed in the browser from two published files at the live platform, where a short dependency-free reference implementation reproduces all thirteen scores. The tag layer is CC BY 4.0; paper text belongs to its authors and is linked, never redistributed.