Detectors Say 'Different', Not 'What': Score, Deviation and Name Are Three Claims With Three Referees, Only the Name Needs a Referee Outside the Detector, and Published Results Show Novelty Detected Far More Reliably Than It Is Named

@misc{mahendrakar2026detectors,
  author       = {Mahendrakar, Pranay},
  title        = {{Detectors Say 'Different', Not 'What': Score, Deviation and Name Are Three Claims With Three Referees, Only the Name Needs a Referee Outside the Detector, and Published Results Show Novelty Detected Far More Reliably Than It Is Named}},
  year         = {2026},
  month        = oct,
  date         = {2026-10-05},
  publisher    = {Life of Research},
  doi          = {10.5281/zenodo.23147856},
  url          = {https://lifeofresearch.com/articles/85},
  abstract     = {(c) 2026 Pranay Mahendrakar. Licensed under CC BY 4.0. An anomaly or novelty detector reports that an input is unlike the data it was built from. Users want more: they want to know what the unusual thing is. Explainable anomaly detection is often presented as supplying that second answer, but three different claims travel under the one word "explanation". A score claim says which parts of the input raised the detector's score; a deviation claim says in what respect the input departs from the reference data; a name claim says what the input is, in a vocabulary a person can act on. This paper argues that the three are checked by different referees: the detector can check the first, the reference data the second, and only something outside both, a label, a benchmark description or a person, can check the third. Assembling published results, it finds that explaining the score is valuable precisely because the score is often driven by something other than the novelty, as generative likelihoods driven by background pixels and an industrial detector partly driven by high-frequency background show; that detectors built on vision-language models do compute their scores from names, yet a correct detection can rest on wrong names; and that where naming has been measured against ground truth it has mostly lagged detection badly: in one open-vocabulary video study, novel categories were detected at 88.2 AUC but named correctly 37.1 percent of the time with the right names in the vocabulary, although a later model on the same task named novel categories on a second dataset almost as well as trained ones. No located study measures, on the same instances, both whether a name is correct and whether it names what the detector responded to. The paper states that missing measurement and the studies that would supply it.},
  keywords     = {anomaly detection, novelty detection, out-of-distribution detection, explainable anomaly detection, open-vocabulary recognition, vision-language models, evaluation methodology, explanation faithfulness},
  language     = {en},
  copyright    = {Creative Commons Attribution 4.0}
}