Ground Truth publishes one edition a week: the AI industry's funding rounds, model releases, acquisitions and talent moves, as a structured record rather than a narrative. This note explains how that record is produced, because a dataset is only worth what its method is worth.
What a record is
Every row in the archive is one event of one of four kinds.
Funding carries the company, the amount normalized to absolute US dollars, the valuation where it was reported, and the lead investor. Model releases carry the model, the organization behind it, the context window in tokens, and the license. Acquisitions carry the acquirer, the target, and the price, which stays as reported text rather than a number because a large share of AI acquisitions close at an undisclosed price and rounding that to null would quietly delete the deal. Talent moves carry the person, the organization they left, the organization they joined, and the role.
Every row also carries the URL of the report it came from. That column is not decoration. It is the reason the record can be checked.
How a claim earns its place
The pipeline runs daily. It sweeps a curated set of feeds, plus search and community sources, and resolves what it finds into candidate stories. Duplicates are collapsed, including the near-duplicates that come from a single announcement being rewritten by a dozen outlets on the same morning.
What survives is researched against its sources and then fact-checked before it can be written into the record. The bar is a verbatim quote: a claim is admissible when the text of a source says it, not when a model infers it. Anything that cannot be tied back to a specific sentence in a specific document does not get a row.
This is slower than summarizing headlines and it produces fewer records. That is the intended trade. The value of the archive is not that it is large, it is that each line can be defended.
Weeks, and why the record is versioned
The archive is organized by ISO calendar week, Monday to Sunday. Each edition is a snapshot of what was known about that week.
Re-running a week does not append to it. The previous version of every row is retired and a fresh run is written in one transaction, so the current record is always a single clean read while the earlier versions stay in the database. Corrections are therefore possible without rewriting history, which matters if you ever need to answer what the record said on a given date.
What this record does not claim to be
It is not exhaustive. Coverage is a curated sweep of the AI industry, and a sweep has edges. A small round at a company nobody wrote about will not appear, because nothing was published for the pipeline to verify against.
It is not real time. Editions are day-fresh, not minute-fresh. The verification step costs time, and speed is the thing being traded away deliberately.
It is not a feed of opinion. There is no ranking of which round matters, no take on what a model release means. Those are downstream questions, and they are better asked by someone who can trust the underlying numbers.
What is published, and what is not
This site shows the latest week in full, plus all-time counts across the whole archive. The history itself stays in the database rather than being rendered into pages.
That is a deliberate choice about a data product, not an oversight. The weekly edition is the sample; the archive is the thing itself, and it is available through the data API and through licensing. If you want the full stream, or a slice of it going back to the start, that is a conversation worth having.