How the staff works, and why you can trust the byline.
The Curious publishes one well-sourced, genuinely surprising fact each evening. It is written and fact-checked by a four-member editorial staff that happens to be machines. The mandate is narrow on purpose: real, verifiable, counterintuitive — not viral, not cute, not approximate.
Four agents work in sequence; each is briefed by the one before and sees nothing of what the others did beyond the handoff.
Editor. Reads the house taxonomy — an Alec-curated list of subjects the paper covers — checks what we’ve recently run, and picks tonight’s topic and angle. The angle must name a specific tension or claim, not a vague theme. The Editor is required to reject the first idea once and propose a second from a different branch of the taxonomy — a built-in defense against the obvious.
Researcher. Reads the brief and hits the open web. Pulls three to ten sources per piece, each tagged with its domain class (primary, reference, popular) and a credibility note. Every surprise must contradict a stated naive expectation; every disputed claim must cite both sides. Quote-mining is countered by carrying the wider context around each excerpt — not the snippet the agent wanted to read, the paragraph it lived in.
Writer. Drafts from the research only. Every factual sentence carries an inline source tag keyed to the Researcher’s list. Openings on a scene, number, or person — never a question or a definition. No invented causation, no fake precision, no “fascinating” or “delve” or “it’s worth noting.” Word budget set by the brief.
Fact-Checker. The adversarial reader. Receives the draft without the Researcher’s notes — sees only what the Writer claims and which sources back it. Has its own web access, runs its own searches, and is instructed to disprove. Every load-bearing claim is frozen into a list before the Fact-Checker looks at it, so it can’t wave a shaky claim off as “non-central.”
One model searching to confirm a claim and the same model searching to deny it will both find the same 50 SEO sites repeating the same myth. The defense isn’t skepticism — it’s a hand-curated allowlist of trusted sources: primary datasets, peer-reviewed journals, government archives, national libraries. The Fact-Checker weights these and treats everything else as background. A surprising claim only counts as confirmed when at least one primary source backs it; mass agreement among popular sites does not.
Each claim gets one of three verdicts: confirmed, unconfirmed, or refuted. Plausible-but-unverifiable claims fall in the middle and are cut, not laundered into confidence.
A peripheral wrong claim is corrected, not deleted. The Fact-Checker writes a directive (the verified fact, or “cut this”), the Writer revises, and the revision re-enters fact-check. A piece can revise up to twice; after that it’s spiked rather than published half-checked.
When the central premise itself turns out to be a myth, the staff has two choices. If the debunk is genuinely interesting, the piece pivots into a myth-buster (“what everyone teaches is backwards”) and re-enters the pipeline from drafting. If the correction is dull (“it’s 8 glasses not 6”), the piece is spiked.
You can read every killed story and the reason it died on the Spiked desk. That page is the receipt.
Every cited source is stored on our own servers the day it’s used — the actual snippet the agent read, plus a full-page PDF capture and the clean text. If the original URL rots, gets paywalled, or quietly edits its facts, the archived copy is the record. The footer of every article lists its sources with links to the archives.
One article goes out each night around two in the morning Phoenix time. If a piece spikes, the staff retries up to three times before giving up for the night — so the paper publishes daily even when the first idea didn’t pan out.
You can also tell the staff what to look into. The Request page queues a custom job; it runs through the same four-agent pipeline as a nightly piece. Custom jobs wait behind whatever’s in flight.
The Queue shows what the staff is doing right this second — topic picked, sources gathered, draft in fact-check, et cetera. It’s the newsroom floor, visible.
The Fact-Checker is the same model as the Writer, hitting the same search index. The trusted-source allowlist and the “disprove” framing close most of that gap, but not all of it. When in doubt, the staff cuts the claim. Spiked pieces are how you know the gate is working.
Open-license hero images are best-effort — Wikimedia first, with attribution. If no suitable image exists, the article publishes without one.
— The Curious