How we make this
Signal Sentry is a twice-daily AI briefing. AI drafts every Drop from sources our bot fetched for that run, and code checks it before it can publish.
When it publishes
Two Drops a day: the Morning Drop and the Evening Drop. Each one stays at its own permanent address in the archive.
Where the news comes from
Our bot reads public feeds and official APIs. It never logs in, never pretends to be a browser, and honors robots.txt.
- Model prices and context windows from the OpenRouter models API.
- New and trending models from the Hugging Face Hub API.
- New and tracked AI repositories from the GitHub API, with star counts from our own snapshots.
- Front-page and recent AI stories from Hacker News, through its search API.
- RSS and Atom feeds from AI labs and news outlets, plus the news pages of labs that have no feed. We show headlines and link to the original. We never republish articles.
- Open roles at AI labs, counted from their public job boards.
- Benchmark results from Epoch AI (CC BY) and the LMArena leaderboard (CC BY 4.0), credited wherever they appear.
- Posts on X, found by Grok (from xAI) and verified as described below.
- Reddit threads, only when they surface through another source and pass Reddit's oEmbed check. We don't collect Reddit directly.
How a Drop is made
- Collect: our bot gathers the stories, numbers and posts for this run and records every URL it fetched in a fetch manifest.
- Read: it fetches the full text of the leading stories, where robots.txt allows, so the editor works from the article and not just the headline.
- Edit: an AI editor, Claude (from Anthropic), picks, ranks and summarizes the stories and writes each "why it matters" line. It works only from what the bot collected. It has no web access and can write nothing but its draft.
- Check: code validates the draft. Every source link must be in this run's manifest, and the checks reject unsafe links, install one-liners, instructions aimed at AI, and hype words. A Drop that fails is not published.
- Publish: the site is rebuilt as static pages, and the build runs the same checks again.
Nothing invented
Every source link on every item must appear in that run's fetch manifest, the record of what our bot actually fetched. Otherwise the build fails, so the editor can't add a link nobody fetched.
Numbers are copied from the collected data. Claims are attributed ("Google says"). Posts on X and Reddit are reported as what people said, not as facts.
What is AI-made
AI writes every Drop: the headline, the 10-second version, the summaries and the "why it matters" lines. The editor is Claude (from Anthropic), working from the sources linked on each item.
Posts from X are found by Grok, xAI's assistant. Grok only points us to posts. Each post must pass X's public oEmbed check: it has to exist and be by the account Grok named. The time encoded in its ID must also fall inside the run's window. Posts that fail are dropped, and Grok's own summaries are never published.
No person reviews each Drop before it publishes. The site owner sets the sources, the rules and the checks, and corrects mistakes in public.
Privacy
No cookies, no accounts, no third-party scripts. Fonts are served from this site.
Your browser stores the last Drop you saw, so new items can be marked, and, if you play a game, your streak. Neither ever leaves your device.
Corrections
When we get something wrong, we fix it in the next Drop and log it here.
We never quietly change the facts in an archived Drop.
No corrections yet.