X releases expanded recommendation algorithm code and transparency tool

X has expanded its open-source disclosures by releasing a significantly larger portion of the code behind its “For You” recommendation feed. The updated codebase, hosted on GitHub under an Apache 2.0 license, includes the “Phoenix” ranking system, model configurations, and the weights assigned to various engagement signals. According to Keith Coleman, X’s Vice President of Product, the new release is roughly 10 to 15 times larger than previous versions, providing insight into how posts are selected, scored, and filtered.

Alongside this code release, X is testing a new “Under the Hood” tool designed to address concerns regarding “shadowbanning” or visibility reduction. The feature allows eligible users—currently a limited test group—to download a JSON file containing aggregated data from the previous month. This file reveals whether specific labels or ranking systems have affected the reach of their posts.

While the transparency initiative provides more detail on content selection, some safety systems remain private. X stated that systems using Grok to assess rule violations will not be disclosed to prevent bad actors from bypassing moderation. The ranking logic reveals that different interactions carry distinct weights; for example, a predicted URL share is weighted significantly higher than a predicted “like,” while predicted reports or mutes act as heavy negative weights to limit visibility.