{
 "row": "model-collapse",
 "address": "model collapse",
 "entity": "model collapse (the concept)",
 "row_type": "C (a conventional reading holds it; the archive extends it)",
 "aliases": [
  {"pattern": "\\bmodel[- ]collapse\\b", "flags": "i"}
 ],
 "classes": [
  {"term": "recursive training on generated data", "pattern": "\\b(recursive(ly)? train\\w*|trained recursively|synthetic data|model-generated (data|content)|generated data)\\b", "flags": "i",
   "basis": "field: B1 Def. 2.1 (F1) 'the data they generate end up polluting the training set of the next generation'; B2 (F12) 'trained on AI-generated content'"},
  {"term": "tail loss", "pattern": "\\b(tails? of the (original )?(content )?distribution|lose[s]? the tails|tail[- ](loss|thinning|mass|pruning)|long[- ]tail)\\b", "flags": "i",
   "basis": "field: B1 Abstract (F2) 'tails of the original content distribution disappear'; B2 (F14) \"'long-tail' ideas might eventually fade\""}
 ],
 "exclude_deposits": [],
 "notes": "Consistency check of D/R/O against Appendix A's S1–S3 selection (27 admitted). Class terms are the field's own phrasings, by locus."
}