How we know

Method

Every article on this site is generated by a pipeline that has to prove its work before anything goes live. This page documents how — including what the system cannot promise.

The claim store

Nothing is written from thin air. Before an article is drafted, its factual material is entered as individual claims in a verification store. Each claim carries three things: a source, a confidence label, and a reference. The writer is only allowed to use what is in the store; a draft with too little verified material is blocked before a single word is generated.

What the confidence labels mean

Sourced
Traced to a named document we have read in full. The reference points to the chapter or section.
Observed
We watched it happen in our own running systems — a logged run, a caught failure, a receipt with a checksum.
Inferred
Our own reasoning built on sourced or observed material. Always labeled as a hypothesis in the article, never presented as fact.
Mythbuster
A limit or counter-claim we keep on purpose — the honest case against our own thesis. Every article is required to carry one.

The gates

A draft passes two independent checks before publishing. First a grounding guard: if the verified material behind the article is too thin, the run stops. Then an independent judge — a different model than the writer — scores the draft on insight, readability, balance and voice. A failing verdict blocks publication outright; the draft is not softened and shipped, it is stopped. This is not theoretical: the judge has blocked our own flagship draft, and the revision only went live after a passing verdict.

Receipts, not assurances

Between the automated steps, every hand-off carries a receipt — a checksum or byte count that proves the work exists as claimed. A green status light proves a step executed. The receipt is what proves it executed correctly. We wrote about the difference in What a Green Checkmark Doesn’t Prove.

What we cannot promise

No honest system can promise hallucination-free content, and we don’t. The judge is itself a model and can be wrong in both directions. Confidence labels are our own assessment, not third-party certification. And we do not yet have infrastructure that measures how often external AI answer engines actually cite us — where that gap matters to an article’s argument, the article says so.

That is the whole method: proof you can read beats confidence you have to trust.