How The Deep Feed uses AI

Drafting: An editorial agent built on Claude Sonnet 4.6 (Anthropic) drafts and copy-edits from primary sources.
Review: A human editor reviews every piece and merges it. The site does not change until that merge.
Banners: Typographic, rendered by a deterministic script. No image model.
Illustrations: Inline illustrations are AI-generated, then scored by a vision model for density and brand.
Byline: One institutional byline, The Deep Feed. No AI system is credited as an author.
Corrections: Errors are corrected in place and recorded on the corrections page.

The pipeline, step by step

SCOUT ranks sources. It collects candidate stories from the sources listed below, removes duplicates and stories the site has already covered, and has a model score each one for newsworthiness, novelty, fit and whether primary sources exist. The best candidates join an idea queue.
DESK drafts. For a news piece, DESK fetches the public primary-source pages attached to the idea and drafts from them. For a long research piece, a supervised run first gathers an evidence bundle. A drafting role writes the piece, a copy-editing role tightens it, the banner script renders the banner, and for research pieces an image role writes illustration prompts.
CRITIC gates the draft. Ten gates run against it. A failure goes back to DESK with the findings, up to three attempts. After the third failure the run stops and is escalated to a person. A high-confidence anonymity finding is escalated too, and stops a scheduled run at once.
A human editor reviews and merges. The agent can at most prepare a draft pull request; it has no merge step. A scheduled run stops earlier still, at a plan for that pull request.
The merge publishes. A merge to the main branch triggers the build and deploy. The deploy re-runs type checks, tests and the full site build before anything goes live.
Distribution waits for a person. A script extracts a social thread scaffold from the published piece without a model. It posts only after a person edits it and marks it ready.

Where SCOUT looks

Hacker News Front page and recent Show HN posts, through the public Algolia API.
GitHub Recently active repositories and releases from ten watched organisations, including Anthropic, OpenAI, Hugging Face and the Model Context Protocol project.
RSS AI sections of The Verge, TechCrunch, MIT Technology Review and VentureBeat.
Google News Through a paid scraping endpoint on the Monid data service. Runs only in supervised sessions where the host supplies that service’s key; held in unattended runs until a spending cap can be enforced.
Exa Paid web search. Held in unattended runs.
X Posts from a watchlist, through the Monid data service. Held in unattended runs.

Which steps use a model

The model names below are the defaults in the pipeline code. The Anthropic calls the editorial agent makes (ranking, drafting, editing and the anonymity judge) each reserve a cost ceiling in a budget ledger before they run, against a default limit of $15 a day, and paid work halts when the limit is reached. Calls that cannot be bounded that way are held in scheduled runs.

Story ranking (SCOUT): Claude Sonnet 4.6, Anthropic API Unattended: Runs. Each call reserves a cost ceiling in the daily budget ledger.
Research (DESK, long pieces): Claude Sonnet 4.6, reading an evidence bundle from a paid research script that includes X posts, LinkedIn commentary and Google News through the Monid data service Unattended: Held. Long research pieces run only under supervision.
Drafting and copy-editing (DESK): Claude Sonnet 4.6, Anthropic API Unattended: Runs for news pieces only, from public primary-source URLs already attached to the idea (at most four). Metered.
Image prompts: Claude Sonnet 4.6, Anthropic API Unattended: Held. News pieces drafted unattended carry no inline images.
Inline illustrations: Gemini 3 Pro Image (Google), through OpenRouter Unattended: Held.
Illustration check at generation: Claude Sonnet 4.5, through OpenRouter Unattended: Held, with the illustration step.
Anonymity judge (CRITIC): Claude Sonnet 4.6, Anthropic API Unattended: Runs. Metered.
Quality judge (CRITIC slop gate): Claude Sonnet 4.5, Anthropic API Unattended: Held. Runs on the pull request and before merge.
Image density check (CRITIC): Gemini 2.5 Flash (Google), through OpenRouter Unattended: Held. Runs on the pull request.
Answer-engine citation probe (CRITIC): Grok 4 Fast (xAI), Perplexity Sonar (through OpenRouter), GPT-4o search preview (OpenAI), Claude Sonnet 4.5 (Anthropic), Gemini 2.5 Flash (Google) Unattended: Held. Runs on the pull request.

Steps not listed here use no model: the banner and social-card renderer, the link, quote and citation checks, the post auditor and the social thread scaffold.

What CRITIC checks

CRITIC is a set of ten blocking gates, one file each. A draft passes only when every gate passes.

anonymity Three passes for traces of private identity: a blocklist scan that also catches spaced-out and look-alike spellings, a style check, and a model reading the draft as a hostile reader trying to identify the author.
blockquote Every block quote must be attributed: a link inside it, or a named speaker in the lines just before it.
citation-floor Counts unique external source links: at least four for a long research piece, at least two for a news piece.
deep-sweep A second pattern sweep for machine-prose habits the main linter under-weights: stock hedges, rhetorical pairs, filler and comma cascades.
external-links Fetches every outbound link. A 404, 410 or persistent 5xx fails the gate.
forbidden-phrases Runs the house post auditor and blocks on any hit in its phrase, construction, verb, adjective and structural lists, plus every pattern added after a past review. Also lints the social thread draft if one exists.
image-density Scores each inline illustration with a vision model. Density below 7 out of 10 fails.
internal-links Every internal link must resolve: post slugs, image files and site routes. Mismatched heading anchors are reported as warnings.
schema-aeo Confirms the structured-data fields the site templates must emit (article metadata, citations, FAQ markup), then runs a probe that asks five answer engines twenty questions and records whether The Deep Feed is cited.
slop A model scores the draft from 0 to 1 against the publication’s best past pieces on primary sourcing, argument structure, specificity, voice and originality. Below 0.7 fails.

When the gates run, and what is held

Inside DESK, after every draft attempt. On every pull request that changes a post, as a CI check supplied with the provider keys. In that check a gate that cannot run its full test fails instead of passing quietly, with one exception: the image density check fails open. A missing key, a scorer error or an unreadable file is logged and does not block. A separate image check on pull requests that add illustrations behaves the same way.

In scheduled, unattended runs three paid checks are held because their calls bypass the budget ledger: the quality judge in the slop gate, the answer-engine probe in schema-aeo (its template checks still run), and the vision scoring in image-density. A held check is noted as held in the gate report, and the editor runs it before merge. The anonymity judge is metered and runs in every mode.

What the editor reviews before merge

The draft against its cited sources: every number, quote and claim traced to the document it came from.
The CRITIC report, including any gate marked as held or skipped. Held checks are run before the merge.
The house post auditor, run on the piece before commit.
The byline and frontmatter: authors: ['the-deep-feed'], no personal names, no AI system credited.
Each inline illustration, by eye, for accuracy and brand.

The review is a human editorial decision. The checks inform it; they do not replace it.

How images are made

Banners and social cards are not AI-generated. A deterministic script sets the headline, category and date in type (rendered with Satori and resvg) on a fixed layout. The same input always produces the same image, and no image model or external API is called.

Inline illustrations are AI-generated. The house image script sends an art-directed prompt to Gemini 3 Pro Image through OpenRouter. A vision model scores each result for density and brand faults, and an image scoring below 7 out of 10 or carrying a fault is regenerated with a revision note, up to three attempts. If the scorer is unavailable the image is kept unscored; if no attempt passes, the best one is kept and flagged for manual review. The image-density gate scores it again before merge.

The byline, and why no AI is credited

Every piece runs under one institutional byline, The Deep Feed. Models are tools in the production process, the way a spell checker or a search engine is. Authorship means responsibility for what is published, and that responsibility sits with the publisher and the editor who merges the piece, not with a model. When an AI lab or model is the subject of a piece, it is named as the subject. The masthead explains the byline policy in full.

Corrections

Errors are corrected in place and recorded on the corrections page. To report one, write to corrections@thedeepfeed.ai with the URL and a primary source.

The EU AI Act, Article 50(4)

Article 50(4) of the EU AI Act, which applies from August 2, 2026, sets disclosure duties for AI-generated text published to inform the public, and treats text that has undergone human review or editorial control, with a person holding editorial responsibility, differently. This page describes the practice; it is not a legal opinion.

In practice: the text on this site is AI-assisted. Every piece passes through human editorial review before publication, and a person decides whether it is merged. Editorial responsibility for everything published sits with the publisher named on the masthead. This page is the standing disclosure of how AI is used.