X Algorithm Simulator

Interactive simulator for the public X recommendation pipeline.

Pinned to xai-org/x-algorithm@590f392 · 简体中文

Overview

The simulator visualizes the published primary default path on deterministic local fixtures. It cannot predict production reach or engagement.

It covers the published source, filter, scoring, and ranking stages, plus fixture-only feed blending. It does not run model training or inference, remote retrieval and hydration services, production ad delivery, or notification filtering.

Screenshots

Screenshots below come from the last full E2E run. More screenshots, the report, and traces are in artifacts/e2e.

Page	Before	Current
Output Explorer
Ranking Simulator

Features

• Step through source, filter, scoring, ranking, and feed-blending stages.
• Expand a ranked post to see up to three weighted contributions, each with an absolute share of at least 10%.
• Inspect 29 Phoenix output heads, ranking weights, filters, visibility, brand safety, and VMRanker DPP.
• Switch the UI between English and Chinese.

Requirements

• Node.js ^20.19.0 || >=22.12.0 and npm, as required by Vite 7.
• ego lite and the ego-browser CLI for the end-to-end tests.

No environment variables, credentials, or external services are required.

Quick Start

git clone https://github.com/lingengyuan/x-algorithm-simulator.git
cd x-algorithm-simulator
npm install
npm run dev

Open http://localhost:5173.

Routes

Path	Page
/	Output Explorer: 29 Phoenix output heads on local fixtures
/simulator	Ranking Simulator: pipeline walkthrough and final feed
/weights	Weight Lab: compare fixture outputs and tune weights
/history	Analysis history stored in the browser

Project Structure

File	Contents
src/data/upstreamSnapshot.ts	Pinned commit and published defaults
src/core/pipeline.ts	Source, filter, scoring, ranking, and feed-blending path
src/core/feed.ts	slot_spread ad blending, brand safety, and feed modules
scripts/e2e.mjs	End-to-end checks, screenshot capture, and trace export

Simulator Scope

• The default ad blender is slot_spread: it reserves spaced ad slots, fills each pair with the nearest compatible posts, and truncates the page.
• Brand-safety verdicts use the legacy path because rust_home_mixer_ads_bs_v2_exp_enabled is false.
• post_video, grok_analyze, and share_via_copy_link weights stay at 0.0, and the video carousel and its CTR model are off.
• Home excursion heads (video continuation and profile visit) are returned with fusion weights 0.0. VMRanker post-topic and recent-view inputs stay off because rust_home_mixer_vm_ranker_send_post_topics is false. DPP-history inputs stay off because rust_home_mixer_vm_ranker_send_dpp_history is false.
• The public repository omits xai_post_text, so local multilingual tokenization may differ from production.

Development and Testing

Run ESLint, TypeScript, the production build, and the ego-browser E2E suite:

npm run check

The E2E runner starts a preview server on port 4173. Results, screenshots, and compressed pipeline traces (.json.gz) go to artifacts/e2e/. Each trace records the input fixtures, the request context, and the weights.

Run the E2E suite alone with npm run test:e2e, or the DOM recovery scenario with npm run test:dom-recovery.

Maintenance

Update the snapshot and implementation together, and pin upstream changes to a specific commit. Keep these behaviors when changing the pipeline:

• The default path reuses Home Mixer weighted scores in VMRanker.
• Topic requests skip DPP.
• Remote retrieval, embeddings, and Thompson draws use repeatable local fixtures.

Daily upstream checks run in the monitor-upstream workflow.

License

MIT