How the Embodied AI Data Ledger is built

A row is a claim about a document. The verdict says how far the ledger got with it. A verified row was fetched at the body that issued it and is quoted from that text. An announced row records that a body said it will act and stops there, because a statement of intent is not an instrument. An absent row records a search that found nothing, and the search itself is written into the record so a reader can repeat it. An open row is a question with no settled answer, which the ledger poses and does not answer.

Supply is who is collecting the hours: state training grounds, company gyms, public datasets. A standard is a document about the data itself, its quality, its format or its collection. A rule is a law that decides what a robot maker may do with hours recorded in a place, whether it may move them across a border and whether the people in them have a say. An open question is where those three run out.

A primary source is the body that made the claim: the regulator's own release, the standards body's catalog entry, the company's own announcement, the law as published in the official journal, the report as issued. Trade press is a lead, never a source. Where a primary document could not be reached, the row says so, the source is labeled secondary, and a verified verdict is not given. A state broadcaster reprinting a research institute's figures is labeled as what it is.

Every figure on the ledger is a number a source printed, in the unit the source used, with the date it was true beside it. The ledger never sums figures from different sources, never converts one unit into another, and never prints a forecast. A legal application date is a fact and appears; an expected publication date is a claim with an expiry and does not.

A check runs against the dataset before any change ships. It fails on an empty dataset, on a record without an https source and a date, on a verified record without a primary source, on an absent record without a written search, on a figure that points at no source or carries no date, on any promised month or year, on an answer under thirty or over eighty words, and on any dash in rendered copy. The check plants nine broken records and must catch every one before it is allowed to look at the real data, because a check that has only ever passed is not evidence.

As of 2026-09-13, 37 records carry 74 primary sources and 47 figures. Those three numbers are counted from the dataset at build time, like every other number on this tool.

Every record is re verified weekly, and the same day for any announcement by a data regulator. A change to a record moves its last modified date, appears in the changelog, and moves the sitemap date for that page alone. Something missing or wrong is a bug, and we want to hear about it.