xai-org/x-algorithm

This system is a recommendation and content classification engine that processes and ranks user-generated content. It combines a modular classification pipeline for safety and spam detection with a retrieval and ranking service that uses transformer-based models and approximate nearest neighbor search. The architecture integrates message queues and data loading services to handle content understanding, ad blending, and engagement prediction.

The repository now includes the 'grox' package, a new open-source component for content understanding and recommendation. This addition introduces a modular architecture for classifying posts and replies, featuring a base ContentClassifier and specific implementations for safety/PTOS checks, spam detection, and reply ranking. The system integrates with Kafka for message ingestion and Strato for data loading, while leveraging LLM-based sampling for classification tasks. It also includes a multimodal embedding engine (V2 and V5) for post and reply embeddings, and a dispatcher/engine layer to orchestrate task generation and execution.

grox · high confidence

This change introduces several new components to the recommendation and home feed pipelines. In the candidate pipeline, a Rust utility for shortening type names is added. In the Grox classifier, new Python modules are added for initial screen content classification (banger\_initial\_screen) and post safety screening (post\_safety\_screen\_deluxe), which use LLMs to analyze and score posts. In the home-mixer, new Rust modules are added for blending ads into the feed (partition\_organic\_blender, safe\_gap\_blender, and their utilities), as well as new candidate hydrators for brand safety, blocked-by status, engagement counts, and filtered topics. Additionally, new Python modules are added for ASR processing and message queue loading, and new types are defined for task scheduling.

candidate-pipeline, home-mixer, python · high confidence

The Phoenix recommendation system is now open-sourced, providing a mini-model (128-dim, 4-layer transformer) trained on real-time engagement data. The release includes a complete end-to-end pipeline (\run\_pipeline.py\) that performs retrieval using approximate nearest neighbor search on a sports corpus, followed by ranking using a per-action engagement model. The artifacts include pre-computed embeddings, model parameters, and example user sequences. Additionally, new test coverage has been added for the attention mask, right-anchored RoPE positions, post-age bucketing, continuous value normalization, and candidate tower configurations.

phoenix · high confidence

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

• CAI 51 → 55 (+4.1)

• Code Health 92 → 92 (-0.0)
• Architecture 100 → 100 (+0.0)
• Maturity 66 → 66 (+0.0)
• Readiness 15 → 24 (+8.9)
• Security 93 → 93 (+0.0)
• Domain Modelling 100 → 100 (+0.0)

• No automated tests
• No tests found

• Coverage not included — suite not readable by the collector

• Unchanged — 0 containers · 1 contexts · 0 edges

Written by watchdog.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

xai-org/x-algorithm was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.