Equity data is a solved problem. Everything around the company isn't.

If you need a company's revenue for the last forty quarters, you have a choice of vendors. Statements, prices, estimates, ownership, transcripts: all of it standardized, keyed to an identifier, delivered by API and fought over on price. For SEC filers the regulator did most of the work by requiring every number to be tagged in XBRL. Equity data is, for most purposes, a solved problem.

Now try a different set of questions about the same company.

• Who sits on the monetary policy committee of the central bank that sets its funding cost, and since when?
• Which bills now moving through the legislature touch its industry, and how far along is each?
• Which vessels called at the port its exports leave from last week, and which chokepoints did they pass?
• How far along is the development it announced two years ago, as seen from orbit rather than from the press release?
• Which organisations does its home country belong to, which agreements has it signed with the countries it sells into, and which conflicts has it been party to?

There is no vendor for any of that. There is a website per source, a PDF per year and a language per country. Each question is answerable, and each answer takes an afternoon.

Why equity data got solved first

Equity data got a schema because a regulator imposed one. Companies have to file, in a prescribed form, on a prescribed calendar, under an identifier that never changes. The buyers were the best-funded desks in finance, so the vendors that standardized it were well paid to do so.

Nothing similar exists for the world around the company. A ministry publishes its current ministers, not who held the post last year. A central bank announces its new governor in a press release, not a table. A parliament's bill tracker, a port authority's schedule and a treaty registry each have their own format and their own terms of use, and none of them shares an identifier with the others or with the company. The history is the part that goes first: most of these sources overwrite themselves.

So the work stayed manual. An analyst carried it in their head: they knew who ran the central bank, they read the news, they remembered the election. That worked well enough when the analyst was the one doing the research.

Agents don't carry it in their heads

Research is moving to agents, and an agent has no memory of last year's election. When it needs the world around a company it does the only thing it can, a web search, and gets back fragments: undated, unsourced, sometimes years stale, and impossible to join to anything else. The answer reads exactly as confidently as one built on filings.

That is the gap we keep running into. A capable model given good fundamentals understands a company. To understand the company's position it also needs the politics, the policy, the geography and the people around it, as data it can query, with dates and sources attached. Today the model is not the limit. The data is.

What it takes to build

Scraping more pages won't close the gap. It needs the same discipline that made equity data usable, applied to sources that were never designed for it:

• Entities, not documents. Countries, governments, central banks, international organisations, people, conflicts, vessels and physical assets, each a record of its own and linked to the companies it affects.
• History kept. Who held a seat, and when; what a series said on the day it was published. A source that overwrites itself has to be captured before it does.
• Every row keeps its source. Each record points back to where it came from, and only reviewed records are served. If a fact can't be traced, an agent shouldn't be repeating it.
• Missing is not zero. A relationship is shown from both sides, and a gap stays a gap instead of being filled with a guess.
• Permission before publication. Much of this information is public without being free to redistribute. We ask the source, and serve what we're allowed to serve.
• Delivered where agents work. Over MCP, the API and the command line, with the terminal for the analyst who steers them.

The layer that's missing

Company data was the first layer, and it is well served. What surrounds the company (the governments, the central banks, the trade routes, the people, the agreements and the conflicts) is the layer nobody sells. That is the one we're building at Astreus.

If there is a dataset your agents keep reaching for and can't find, tell us. Those requests decide what we build next.