Editorial policy

Guessarium makes factual claims about history, and it publishes them next to a scoreboard. A wrong date on a quiz answer is a small thing. A wrong date on a page a teacher hands to a class is not. So the write-ups get treated like reference copy, not game flavour text.

One. The draft. An AI model writes a one-sentence summary and roughly a hundred words of context for the event. It's told to stay general rather than guess when it isn't sure of a figure or a date.

Two. The gate. An automated check scores the draft out of a hundred across three areas: whether the facts hold up, whether the writing reads like a person wrote it, and whether the card is structurally complete. A card needs 85 or better and zero blocking problems. A card that fails gets sent back for one rewrite with the specific problems attached.

Three. The independent audit. A different and stronger model then audits the finished draft, in several passes. This matters more than it sounds. The gate's fact check is the same model family that wrote the card, and a model grading its own work is a weak check. We measured it: over one catalogue sweep the independent auditor caught eight cards the gate had passed, five of them at a perfect score, including a garbled chronology and a migration described in the wrong direction.

Four. The source check. When the audit flags a claim as serious, that claim gets checked against real sources on the web, and the sources get ranked by what they actually are. A reference work or a primary record counts as evidence. A forum post or a blog does not, and a card is never marked verified on one.

Five. The human. A person reads the card, with both audits and every finding in front of them, and approves it or sends it back. Approval is the act that lets a page be indexed. Until it happens the page carries a noindex tag and search engines are told to leave it out.

They disagree often, and that's the point. The two auditors find different things, so neither one replaces the other and both verdicts are kept. The disagreements are what the human reviewer looks at first.

Corrections aren't applied blindly either. When a finding looks legitimate, a stronger model evaluates it as a suggestion, applies the ones that are right, and rejects the ones that aren't. It has refused fixes that would have made a card less accurate. Whatever survives is audited again before it's kept.

Review is attributed to the Guessarium editorial team. We don't put a personal byline on these cards, and we don't claim credentials the work doesn't rest on.

These are short reference cards, not scholarship. Some events are genuinely contested, and where the history is disputed the card should say so rather than pick a side quietly. If you find something wrong, tell us and we'll check it. Corrections change the page's updated date. Nothing else does, because a freshness date that moves without the content moving is a lie told to a crawler.