Editorial Standards

What we publish

We publish two kinds of work. Atlas entries, 6,543 of them, each place a food, dish or moment in a particular place and time. Each has a written reconstruction, historical context, an evidence assessment and an image. Research covers investigations, curated collections and answered questions. The editors write it, it is built on the atlas, and it goes through the stricter review described below.

Sources

We prefer, in this order: excavation and site reports; peer-reviewed scientific literature; primary sources from the period, such as texts, records and recipes; and recent scholarly syntheses. A source is listed only when it supports the claim it accompanies. Each citation is checked twice: once to confirm the publication exists (its DOI must resolve in Crossref) and once to confirm it says what we use it for.

Historical entries list their sources now. Dish entries gain theirs as the sourcing work reaches them. In research pieces, an atlas entry is a lead, never evidence: a piece cannot cite the atlas as proof. We link to publishers instead of copying their text, and any quotation is short and attributed.

Evidence labels

Every entry rates its evidence one category at a time (dating, food identification, preparation method and so on), using one of these labels:

• Strong: directly attested, by remains, a dated deposit, or a contemporary text that names it.
• Moderate: well supported, but indirectly or from a nearby site or period.
• Probable: the most likely reading of the evidence, with credible alternatives.
• Limited: little direct evidence; the claim rests on analogy or a few data points.
• Interpretive: a scholarly reading or reconstruction, not an observation.
• Insufficient evidence: the record cannot answer the question, and we say so instead of guessing.

Each entry also has a Limitations section. It publishes problems with the entry itself, such as a date that later research has moved or a doubtful attribution, even when that contradicts the rest of the page.

How AI is used

AI systems do much of the work. For atlas entries, they search the literature, draft the text and evidence assessments, propose sources, estimate nutrition and generate the images. They follow written rules: never present inference as observation, never invent a citation, always state uncertainty. Their output passes automated checks and editorial review before it is published.

Research pieces are different. An editor sets the question, verifies the sources and is accountable for every claim. AI may help organise and draft from material the editor has verified. It never supplies a source or a fact that an editor has not checked, and nothing it writes is published without an editor’s approval.

Images are reconstructions

Every scene is generated by AI. None is a photograph or an artefact. Each scene is built from its entry’s evidence: the setting, people, food and tools come from the best-documented material for that place and date. Where the record is silent, the scene’s choices are consistent with what the record does show. A scene shows what the evidence makes plausible. It is never evidence itself; the entry’s Evidence section shows what is known and what is inferred.

How research is reviewed

Every investigation, collection and answered question goes through five steps:

1. A precise question.
2. Research and an evidence table recording each central claim, its source, its strength and anything that conflicts with it.
3. A draft.
4. A fact-check by an editor who did not write it. Kirill Khroustalev checks scientific and evidential claims.
5. Final approval by Alexei Stoliartchouk. When he wrote the piece, another editor approves it. No one approves their own work.

Every published piece names its author and reviewer and shows its dates and sources.

Corrections

Every entry asks “Spot something wrong in this reconstruction?” and offers a way to tell us. An editor reads each report. When a correction changes a fact, date or attribution, we change the record. For research pieces, we also date the correction on the piece and log it on the Research Updates page. We would rather correct than defend.