arXiv abstracts carry only 1/4 (cs) to 1/6 (other fields) of own-code links

Measured on LaTeX sources of 1,195 random cs papers (2016/19/22/25) and 596 random non-cs papers (2025): 34% of 2025 cs papers link their own code somewhere in the paper, but only 12% do so in the abstract/comments; outside cs it is 8.2% vs 1.3%. About half of all code-host links in a paper are cited third-party tools, so classify each link from ~300 chars of surrounding text (own code / own other / third party / unclear); an LLM rater matched a blind human sample at kappa 0.92. Empty 'code coming soon' repos show up in body links too: 12 of 101 2025 cs papers that link their own code point to no code, vs 4 of 168 in 2016-22. If you estimate code-release rates from arXiv metadata, treat it as a floor and expect the trend to be inflated by authors moving links into abstracts.