Editorial Policy
Effective date: 2026-05-21
This page documents how DeepSignal chooses what to surface. It is deliberately specific. If you ever wonder “why did this story end up in Featured?”, the answer is here and on the article card itself.
1. Source selection
DeepSignal monitors sources that satisfy all of:
- They publish primary AI material — frontier-lab announcements, model releases, peer-reviewed research, infrastructure / chip news, well-sourced business reporting, or first-person posts from people shipping the work.
- They expose a public feed (RSS, Atom, JSON, or an equivalent public endpoint) so ingestion is honest and auditable.
- They are referenced repeatedly by other primary sources in our tracking window — a one-off blog post is not a candidate.
The current roster is 27 sources spanning frontier labs (OpenAI, Anthropic, Google DeepMind, Meta AI, Mistral, Perplexity, Hugging Face), independent voices on X (Karpathy, Altman, OpenAI, Anthropic, Cursor), research feeds (arXiv cs.AI / cs.CL / cs.CV, Stanford HAI), community discussion (Hacker News), business and tech press (TechCrunch, CNBC Tech, MIT Technology Review, Fortune, Yahoo Finance, Tom's Hardware, Robotics Tomorrow, The Hacker News), and tooling (Vercel). The full list with category and homepage URL is on the Source Attribution Policy page. The live, user-toggleable version is at /sources.
2. The Signal Score — the actual formula
Every article that enters the pipeline gets a Signal Score from 0–100. The score is the weighted sum of five factors:
| Factor | Weight | What it captures |
|---|---|---|
| Source authority | 30% | Priority assigned to the publisher (OpenAI Blog and DeepMind sit at the top; community feeds like Hacker News and arXiv search are mid-tier). |
| Technical impact | 25% | Breadth of technical tags hit by the article (LLM, Agent, AI Coding, Inference, Robotics, GPU, Open Source, Security). |
| Business impact | 15% | Business tag hits (Funding, Acquisition, AI Startup, Enterprise AI, Policy) plus title keyword detection for funding rounds, acquisitions, IPOs. |
| Novelty (recency) | 30% | Time decay with a 24-hour half-life. A story is at 100% novelty when ingested; ~50% one day later, ~25% two days later, etc. |
| Community heat | 0% | Computed for telemetry (HN points, GitHub stars equivalent) but currently weighted at 0 until reliable upstream counters are wired. |
Source code reference: lib/pipeline/signal.ts. The weights above are the runtime values; they were rebalanced in May 2026 to remove the previously dead community-heat weight and lift premier-source ceilings. If we change them, this page changes with them.
Two additional adjustments are applied at query time:
- Recency decay at SELECT time. Featured ranking uses a live time-decay multiplier on top of the stored score, so a top story slides down the feed as it ages without us having to re-process it.
- Featured threshold. An article is flagged as Featured if its Signal Score is at or above the current threshold (today: 60). The threshold is editable and is logged whenever it changes.
3. AI summaries
Every published article carries an AI-generated short summary, a bullet list, and a one-paragraph “Why Featured” rationale. These are generated by OpenAI's GPT-4o-mini against prompt templates that live in the admin console at ai_prompts and are documented as plain English templates.
The prompt strategy is:
- Provide the model with the full extracted article text plus structured metadata (source, tags, published date).
- Ask for one neutral sentence (the summary), 3–5 short bullets, and a one-paragraph rationale that explicitly justifies why a reader should care.
- Bilingual EN / 中文 outputs are produced in the same pass so the two language versions stay consistent.
- Summaries are reviewed for accuracy issues before publishing, and obvious hallucinations or off-topic outputs are dropped or regenerated.
AI summaries are best-effort, not authoritative. The original article is always one click away — every card and detail page links to it.
4. Featured vs Latest vs Daily Brief
- Latest is the firehose. Every published article appears here, sorted reverse-chronologically.
- Featured is the curated layer. An article must score at or above the Featured threshold (currently 60). Featured also caps per-source representation so one publisher cannot dominate the page.
- Daily Brief is generated each morning. It pulls the highest-signal stories from the prior 24 hours and groups them by vertical (models, agents, infra, hardware, robotics, security, policy, business). Vertical assignments come from the tags that drove the Signal Score, so the daily brief inherits the same editorial logic.
5. Corrections and takedowns
If we publish something inaccurate, we correct it. Email contact@ai-deep-signal.com with the URL and the specific error; corrections are typically applied within 24 hours, and a brief note is added to the article if the change is substantive.
Publisher takedown requests are honoured within 48 hours from good-faith requests, per the Source Attribution Policy.
6. Independence
DeepSignal does not accept payment to feature stories, raise scores, or include sources. The editorial pipeline and the Signal Score formula are the only thing deciding what shows up. If that ever changes — for example, through a sponsorship or partnership — this page will say so before it happens.