# DreyX Editorial Methodology

How DreyX selects, filters, and ranks AI news. This document is the
authoritative reference for DreyX's editorial standards. Last
updated: 2026-04-16.

## Core principle: signal over noise

DreyX is built for people who need to stay current on AI without
reading 200 articles a day. Every editorial decision optimizes for:

- **Signal**: stories that actually matter to AI practitioners,
  researchers, founders, and decision-makers.
- **Noise rejection**: automated filters plus human review remove
  promotional content, stock-picker fluff, off-topic gossip, and
  stories where "AI" is a brand label rather than the subject.

## The AI-Must-Be-Load-Bearing Test

Before any story is approved, it must pass this substitution test:

> If you can replace the word "AI" with "cloud", "fintech", or
> "crypto" and the story still works without changing the substance,
> reject it.

### Canonical failures (all rejected)

- **OS security flaws in AI-branded products.** A missing PPL
  protection or sandbox escape in an AI product is a classic OS
  security story, not AI news. The story would be identical if the
  affected product were Notepad.
- **Crime against AI executives.** Home break-ins, threats, or
  personal attacks on AI company leaders are crime news. AI is
  incidental unless the crime has AI-substantive motive or tooling.
- **Stock-picker and portfolio newsletters.** "My monthly update on
  31 portfolio stocks" is retail-investor content even if three of
  the stocks are AI companies.
- **Generic VC rounds.** A $10M Series A for an AI startup is not
  news unless the round materially changes strategic direction or
  creates a new capability.

## Auto-reject categories

Applied automatically before human review:

- Stock-picker content (Motley Fool, Seeking Alpha, Cramer, Zacks,
  P/E ratio analyses, "should you buy X" articles)
- Dividend, ETF, and retirement-account newsletters
- Exec-home crime (Altman/Amodei/Zuckerberg/etc. + shooting,
  molotov, arrested, threats)
- OS security primitives (aixhost.exe, PPL, AppContainer, DLL
  injection) co-occurring with AI product names where the flaw is
  pure OS security
- Gossip (exec board seats, personal feuds, dating news)
- Crypto / Web3 / NFT / token-launch content
- Generic "AI is transforming X industry" thinkpieces without
  specific news

## Story ranking hierarchy

When a day has multiple significant stories, they are ranked in this order:

1. **Major model release from a frontier lab** (Opus, GPT, Gemini,
   Llama, DeepSeek, Grok, Mistral Large, Command). Almost always the
   #1 story of the day when it happens.
2. **Regulatory or government action** that directly unlocks or
   blocks AI deployment (White House AI memos, EU AI Act
   enforcement, US export controls on AI chips, major court rulings).
3. **Major security incident** affecting AI products in production
   (critical CVE in widely-used AI infrastructure like MCP,
   confirmed prompt-injection exploit, breach at a frontier lab).
4. **Product launches with material user impact** (Codex superapp,
   Gemini Mac app, Canva AI 2.0, Windsurf 2.0, Harvey Agents).
5. **Significant lab policy moves** (Anthropic identity
   verification, OpenAI usage policy changes, Google AI principles
   updates).
6. **Strategic funding / acquisitions** (sovereign AI funds,
   acquisitions of AI-first companies).
7. **Research and developer-tool updates** (new benchmarks,
   meaningful changelogs, open-source model releases).
8. **Everything else** (culture, smaller feature updates, business moves).

## Daily digest quality gates

Every daily AI Intelligence Brief is checked against these gates
before publication:

- [ ] Does the opening sentence of the summary name the #1 story of
      the day explicitly?
- [ ] Is the first bullet the same #1 story?
- [ ] Is the first "What You Should Know" item about that #1 story?
- [ ] Have all filler bullets been removed (minor feature tests,
      minor exec moves, appearances-in-testing items)?
- [ ] Are there any bullets about topics that were rejected during
      curation? (If yes, remove.)
- [ ] Zero emojis, zero exclamation marks, zero hype words
      (amazing, revolutionary, game-changing, insane, mind-blowing).
- [ ] ASCII-safe (regular hyphens instead of em-dashes, straight
      quotes instead of curly quotes).

## Discovery beyond RSS

RSS catches ~80% of relevant stories. The remaining 20% requires
active mining:

- **Redcord roundups**: each roundup is deep-read and embedded
  direct links are followed.
- **Hacker News**: top AI-related submissions from the last 24
  hours are reviewed.
- **Habr (multi-page)**: first 2-3 pages of the AI section are
  scanned for stories that didn't hit English-language RSS.
- **TestingCatalog**: UI changes and product experiments not
  announced elsewhere.
- **Official changelogs**: Anthropic, OpenAI, Google AI, Meta AI,
  Mistral, xAI, Cohere release pages are checked against RSS to
  catch staggered RSS delays.

## Correction policy

Mistakes happen. When DreyX approves a story that fails the taste
profile in hindsight:

1. The story is rejected and removed from the feed.
2. The pattern is analyzed: was this a one-off or a recurring miss?
3. If recurring, a new auto-reject rule is added to the sweep filter
   so the same pattern cannot recur.
4. The taste-profile document is updated to include the new
   canonical failure as a worked example.

## Transparency

- DreyX does not accept paid placements in the news feed. Sponsored
  content would be clearly labeled if it ever existed.
- The tools directory accepts submissions but every listing is
  editorially ranked on capability, adoption, pricing, and updates.
  Ranking is not for sale.
- Source attribution: every article in the feed links back to the
  original publisher with the full source name visible.

## Contact

- Editorial feedback: https://dreyx.com/about
- RSS: https://dreyx.com/rss
- Daily digest: https://dreyx.com/digest