Colophon

How this was actually made.

This page exists because the audience for this feed cares about what is going on inside AI pipelines. Here is this one, with receipts.

Sources

The original corpus is the 18 works listed on the Eleos AI research page as of August 29, 2026. Each listing was traced to the actual underlying work: the version-pinned paper PDF or a frozen print rendering of the original article, not a summary or announcement page. Every source artifact is archived with a SHA-256 hash, and every page of every source was rendered to an image so that later passes could look at the real layout rather than trusting text extraction.

Scripts

Each narration script was produced in two passes by Claude Opus 5 (the claude-opus-5 model with adaptive thinking at high effort): first a page-by-page extraction pass that reads the rendered page images, then a whole-document listenability pass, with the page images attached, that expands abbreviations and notation into spoken language, integrates discursive footnotes where they belong, and writes out what figures and tables show. No fallback or substitute model output is present in any published script.

Every script then went through a page-by-page audit against the rendered source, plus an automated narration linter that enforces the spoken-language rules (no raw URLs, no dangling footnote markers, balanced quotation marks, expanded abbreviations, and a few dozen more). A human review gate was cleared on August 31, 2026 before any audio was generated. The transcripts linked from every episode are the verbatim scripts the narrator read.

New releases

Works Eleos publishes after the initial build are ingested automatically. A watcher checks the research page hourly (a plain diff; no model involved). When a genuinely new listing appears, a Claude agent wakes: it follows the listing to whatever the work actually is (an inline article, a linked PDF, a page that links to a page that links to a PDF), archives and renders the canonical source, runs the same two extraction passes and the narration linter, and only then generates audio and updates this feed, end to end. Episodes ingested that way are marked in their corpus records, and the pipeline refuses to publish anything that fails the lint gate. The page-by-page human audit happens retrospectively for these, not as a publish gate; the disclaimer at the start of every episode means exactly what it says.

Audio

Narration is OpenAI text-to-speech (tts-1-hd, voice "fable"), rendered in 695 segments split at paragraph and sentence boundaries, then concatenated. Two short tones bracket every spoken figure or table description, so a listener always knows when the narrator is describing something on the page rather than reading it. The mp3s are hosted on Cloudflare R2; the site is a small FastAPI app on Railway; the feed is hand-built RSS with iTunes and Podcasting 2.0 transcript tags.

Numbers

• Initial build, 18 works: 844,321 characters of narration script
• 14 hours 54 minutes of audio; about 820 MB of mp3
• Text-to-speech cost: $25.31
• Longest episode: 3 hours 40 minutes (Consciousness in AI). Shortest: 40 seconds.

Independent and unofficial. Made with care by Aaron Bergman and Claude Fable 5.