Editorial Standards

How Culturily verifies what it publishes

Culturily publishes one edition per culture per day. Every fact in that edition is drawn from a named source and checked by machine before publication. This page describes that process in full, including its limits.

The division of labour

A human selects the sources. Machines verify the facts against them.

Source selection is not automated, and will not be. Judging whether an institution is real, whether a scholar is credible, whether an archive holds what it claims — this is knowledge of the world, not of text. Systems that rank sources by reputation measure how the written record treats a source, never whether it merits that treatment. That reasoning is circular, and it fails precisely where authority is contested.

Sources are admitted only under a written standard: they must be primary, institutional, academic, or a reference work of record with a named editorial board. Aggregators, content farms, and unsourced encyclopaedic mirrors are excluded by name. The test we apply is whether a careful scholar could cite the work in a footnote without embarrassment.

Everything downstream of that judgment is verified by machine, and the verification is the substance of this page.

What every fact must survive before publication

1. It must be traceable to the source, word for word. Each fact is bound to a verbatim passage from a named work. Every date, number, and proper name appearing in the fact must appear in that passage — verified by deterministic check, not by a model’s assessment of its own output. A fact that introduces a detail the source did not state is rejected. It is not corrected, softened, or published with a caveat.
2. It must survive an independent challenge. A second model is shown the claim alone — without the source, without the citation, without knowing what checked it before. Its task is to refute the claim from what it knows. It cannot defer to authority, because it is never shown any. Refutation holds the fact.
3. It must still be true. Claims that can decay — records, rankings, superlatives, present-tense statistics — are either anchored to a stated date, or they do not publish. “Czechia leads Europe in beer consumption” was true when written and may not be true now. “In 2023, Czechs drank more beer per capita than any nation” is a claim about the record, and remains true permanently. We publish the second form.
4. Certain claims never publish automatically. Assertions concerning living persons, medicine, contested politics, religious doctrine, or the character of any ethnic group are withheld from automatic publication without exception. They escalate to human decision or they do not run.

What we do not do

We do not publish what we cannot verify. Where the system is uncertain, it withholds. A thin edition is a supply problem: visible, measurable, and fixable. A false fact is a truth problem: invisible until it does damage. We do not trade the second for the first.

We do not amend the archive silently. When a published fact is found to be wrong, the archive records the correction. This convention is older than we are and we see no reason to depart from it.

Accountability

We select and tier every source, and we are accountable for every published fact. The verification system described above is enforced in code, not in policy: a fact that does not carry a passing verdict from each stage cannot be assembled into an edition. There is no manual override that publishes an unverified claim.

Source selection is performed by a named individual, not a committee and not a model. Culturily is currently a small operation.

Our verification has found errors in our own work. We have withdrawn facts for misattributing a letter to the wrong recipient, for dating a consecration to the wrong month, and for describing a man as Emperor thirteen years before he held the title. None of these reached a reader. They were caught by the process described on this page.

The limits of this process

We state these plainly because a standard that names no limits is a marketing document.

Verification confirms that a claim is faithfully drawn from an authoritative source and survives independent challenge. It cannot confirm that a source is right. Where a scholarly consensus is mistaken, and every authority repeats the error, no amount of automated checking will detect it. This limit is shared by every human editor who has ever worked from the record, and we do not claim to have escaped it.

We publish corrections when we learn of them. If you believe a fact is wrong, every edition carries a means to report it.