Editorial planning
Keeping facts aligned across languages: the RankAI method
A stable ID per fact, human approval before publishing, and a weekly count of rows marked Needs review. How RankAI keeps the same fact consistent across Italian, English, Spanish and German.
In RankAI, every fact carries a stable ID and every version is approved by a person before it goes out. That is the whole rule. Take fact F-014. If the Italian version shows “Needs review” because “by default” is missing, the English version has to show the same value, without “by default”, and the same “Needs review” outcome. If the English value does contain “by default”, then the correct outcome is “Aligned” in all four languages. Human approval of a version is the default. Automatic publishing is switched on per site and still goes through the checks.
Why the stable ID holds the method together
A fact is not a sentence. It is a unit with a code, a value and an outcome. Without a stable ID, the same piece of information translated into Italian, English, Spanish and German turns into four disconnected objects, and comparing versions comes down to reading them by hand. With the ID, the comparison is mechanical: same code, same expected value, same outcome. This is RankAI’s working proposal, not a universal standard.
When the source fact changes after translation
Sometimes the source fact changes after the translations have already been approved. The ID stays the same, but the expected value does not. Every translated version goes back to “Needs review” until a person confirms the new value in each language. Skip that step and a version can look aligned simply because nobody compared it against the updated value.
The weekly count of “Needs review” rows
A count of “Needs review” rows is useless if it stays a number in someone’s head. Log it in a shared register, one row per fact and per language, updated every Friday. If the total for a site goes above the threshold agreed for that site, switch off automatic publishing for that site until things are aligned again. The threshold is the team’s call, but it has to be written down before the count, not after.
Steps that make alignment checkable
- Give every fact a stable ID and carry it into every language. Check: the same code appears in every version.
- Compare value and outcome by ID, not by sentence. Check: row F-014 has the same value and the same outcome in Italian, English, Spanish and German.
- Every Friday, log the “Needs review” rows in a shared register. Check: the register shows date, site, ID and language.
- If the count passes the written threshold, switch off automatic publishing for that site. Check: the register shows the switch as off.
What this piece adds to a plain comparison table is the split between aligning the value and aligning the outcome. Two versions can both show “Aligned” while holding different values, if the comparison was made on the sentence rather than the ID. So we propose, as the RankAI method and not as a universal standard, to record three fields per fact: code, expected value and outcome. The expected value is what a person approved; the outcome is the result of the comparison. If the expected value changes, the outcome cannot stay “Aligned” until a new human approval comes in. In the F-014 case, the Spanish version keeps “página de approfondimento” in Italian: that is a leftover from translation, but the reference to the fact is untouched, so the outcome has to be decided on the fact’s value, not on the language of the link. Keeping these two levels apart stops a translation error from being mistaken for a fact mismatch, and the other way round.
Where the method stops being tested
In the technical trial the limits are 3 articles per month and 3 in total. Past that point the method has not been tried. For the editorial plan and the approval rules, see the detail page.