A paper submitted on 12 November 2026 reports that X’s public recommender code leaves important parts of its production pipeline difficult to inspect. Authors Luka Bekavac and Simon Mayer identify 13 “transparency dead-ends” on the live serving path: models, services, label lists and configuration values that the code depends on but the release does not contain. The paper is under peer review, according to its arXiv record.[1]
The authors describe a code-level audit that walks backward from the user-facing timeline to the upstream components behind each decision. They check whether components are reachable from the served pipeline and classify them using four inspectability tiers. The abstract says the components that decide which posts are filtered and how they are ranked fall in the two least inspectable tiers.[1]
The authors also report that the 2025 integration of the proprietary Grok language model into moderation, annotation and ranking intensifies this pattern. They say the release retains 2023 code that the 2025 pipeline no longer reaches; without a reachability check, 36 parameter references in that code could be mistaken for dead-ends. The paper discusses additional artifacts needed for meaningful auditing, but the saved abstract does not enumerate them.[1]
Lead photo: Luka Bekavac, Simon Mayer · CC BY 4.0.
Sources
Check or correct this page
Verification record
The digest below shows whether the text you are reading matches the published source. It does not show that the reporting is correct. The build computes it from this page's Markdown source and lists it in the site manifest.
a13c35480c122b15bb6b8d1aab8de473d582ee8be167b5948b7eaf39f4a7fdc7
