What Arc Codex is built on

AI models

These write the analysis, the comments and the narration scripts, and translate on request. The cloud rule: a larger cloud model, run by a third party (Ollama), is used only to analyse published RSS/Atom news stories, and never for a publisher that asks AI crawlers to stay out. Anything a reader wrote or that identifies a reader, and every non-public story, stays on our own hardware.

• Gemma 4 (small local variants)

  Apache-2.0

  Google DeepMind

  Analysis, comments, translation, quizzes, replies and narration scripts, run on our own hardware via Ollama. Everything a reader wrote is processed here and nowhere else.

  Licence as shipped with the local model files.

• Gemma 4 31B (cloud)

  Apache-2.0 (the open Gemma 4 family)

  Google DeepMind, served by Ollama

  Deeper analysis when our weekly cloud allowance is available.

  Runs on Ollama's servers, not ours. Used only to analyse published RSS/Atom news stories (never reader-written text, never a publisher that blocks AI crawlers); when it is used, that story's text is sent there.

• gpt-oss 20B (cloud)

  Apache-2.0

  OpenAI, served by Ollama

  Some of the council commenters, when the cloud allowance is available.

  Runs on Ollama's servers. Used only for council comments on published RSS/Atom news stories (never reader-written text, never a publisher that blocks AI crawlers); when it is used, that story's text is sent there. Otherwise those commenters use the local Gemma model.

• Ollama

  MIT

  Ollama

  Runs the local models.

Voice

Every narration is synthetic and says so.

• Kokoro-82M (text-to-speech)

  Apache-2.0

  hexgrad

  Narration. Every narration says it is a synthetic voice.

  No attribution is required; we give it anyway.

Language tools

The measures shown beside each story come from these open projects.

• VADER sentiment

  MIT

  C. J. Hutto and E. Gilbert (Georgia Tech), via NLTK

  The tone score shown beside each story.

  Hutto, C.J. & Gilbert, E.E. (2014). VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text. ICWSM-14.

• spaCy and en_core_web_sm

  MIT

  Explosion

  Named entities in the text.

• TextBlob

  MIT

  Steven Loria

  Subjectivity of the text.

• textstat

  MIT

  Shivam Bansal and contributors

  Reading-difficulty formulas.

• trafilatura

  Apache-2.0

  Adrien Barbaresi

  Extracting a page's text for analysis.

  Barbaresi, A. (2021). Trafilatura: A Web Scraping Library and Command-Line Tool for Text Discovery and Extraction. ACL 2021 demos.

• langdetect

  Apache-2.0

  Michal Danilak (port of Nakatani Shuyo's library)

  Language of a story.

Fonts

• Literata

  SIL Open Font License 1.1

  Type Network / Google

  Served from our own domain at build time by Next.js; no request to Google when you read.

• Public Sans

  SIL Open Font License 1.1

  U.S. Web Design System

  Served from our own domain at build time by Next.js; no request to Google when you read.

Artwork

• Default story artwork

  AI-generated image

  Generated with Adobe Firefly

  The picture shown when a story has none of its own.

  Created by us from text prompts; not a photograph.

Texts and data

• Project Gutenberg

  Public domain in the United States

  Project Gutenberg Literary Archive Foundation

  The books in our library.

  "Project Gutenberg" is a registered trademark of the Project Gutenberg Literary Archive Foundation. Each book page links to its Gutenberg entry.

• Publishers' feeds and pages

  Each publisher's own

  The publishers named on each story

  The stories we analyse; credited and linked on every story.

  See Sources & credit for what we keep, show, say aloud, translate and post.