AVWBeta
ajad van wyk
About
Portfolio
Tech Stack
Blog
Gallery
Stimming
Resources
AVW
AVW
Get in touch

Let'sbuildsomethingremarkabletogether.

Start with the background, selected work, or the profile brief for a sharper read on the practice and the work behind it.

Get in TouchProfile Brief

Explore

About
Portfolio
Blog

Expertise

Finance & Investment
Legal Framework
AI & Innovation
Research & Writing

Resources

Browse Study Notes
Download LLM Profile
Tech Stack
Gallery

Trust

Email
Download Profile Brief
About AVW
Selected Work

Latest Notes

Archive

Phoenix

X Algorithm

Thunder

X Algorithm

X Algorithm — For You Feed

X Algorithm

AVWCape Town, South Africa

© 2026 Ajad Van Wyk. All rights reserved.

Back to Resources
X Algorithm

XAlgorithm—Pre-ScoringFilters

15 Apr 2026
2 min read
recommendation-systemx-twitterfilterspipeline
In this note
  1. 01Why filter before scoring?
  2. 02Filter stack
  3. 03Post-selection filters
  4. 04Execution model
  5. 05Related notes

X Algorithm — Pre-Scoring Filters

"10 boolean filters applied after candidate hydration, before any ML inference. Cheap to run; protect the expensive transformer from wasted compute."
01

Why filter before scoring?

The Phoenix transformer requires a full Grok forward pass per candidate — significant compute. Boolean filters are microsecond-level lookups against preloaded data (block lists, mute lists, seen history). Running them first means the transformer only processes genuinely eligible candidates.

02

Filter stack

Group 1 — Content & recency

FilterRemoves
Age filterPosts older than the retention window
Drop duplicatesDuplicate post IDs
Repost deduplicationMultiple reposts of the same underlying post
Self-post filterThe requesting user's own posts
Core data hydration filterPosts that failed to fetch metadata during hydration

Group 2 — Social graph & user state

FilterRemoves
Author social graph filterPosts from blocked or muted authors
Muted keyword filterPosts containing words in the user's muted list
Previously seen posts filterPosts the user has historically engaged with
Previously served posts filterPosts already shown in the current session
Ineligible subscription filterPaywalled content the user lacks access to
03

Post-selection filters

Two additional filters run after scoring and selection:

FilterRemoves
VF filterDeleted, spam, violence, gore content
Dedup conversation filterMultiple branches of the same conversation thread
04

Execution model

Filters are implemented as Filter traits in the candidate-pipeline crate. They:

  • - Run as part of the Home Mixer pipeline
  • - Have access to user context from query hydration
  • - Can fail gracefully with configurable error handling
05

Related notes

  • - X Algorithm — For You Feed — where filters sit in the pipeline
  • - Phoenix — the ML inference stage that follows filtering
  • - X Algorithm — Key Concepts — query hydration (supplies block/mute lists)
Previous

X Algorithm — Key Concepts

Next

X Algorithm — Scoring Pipeline

More from this module

Phoenix

recommendation-systemx-twitterout-of-network-retrievaltransformer

Thunder

recommendation-systemx-twitterin-network-retrievalrust

X Algorithm — For You Feed

recommendation-systemmachine-learningx-twitterstudy-guide

X Algorithm — Key Concepts

recommendation-systemx-twitterconceptsglossary
Keep exploring

More notes and resources

Browse the full resource vault, or reach out to discuss any of these topics.

All resources
Get in touch