What sight means

By 1570 the maps of the world already existed. Abraham Ortelius did not draw new ones. He collected the maps other people had made, redrew them to one standard and bound them in one book. With the whole world in front of him he noticed that the coasts of the Americas and Africa looked like they had once fit together. He was right, and it took science about three hundred years to catch up. Nobody had sailed anywhere new. The facts were all known. They had never been side by side.

That is what I mean by sight, and it is the real payoff of a world model for a company. Getting facts right is where the model earns its keep first, and we can show that it does. Seeing what no single source contains is the larger bet. I want to say plainly that it is the one we are making.

The company version is a regulation next to the accounts it touches

A regulation changes. In most companies that fact lands in one place: a newsletter someone forwards, a remark on a call, a line in a Slack channel. The people it affects are somewhere else. So are the deals. In a world model the regulation is an object, and it is linked to the facilities it applies to, the companies that run them, the open opportunities with those companies, the earlier calls where it came up and the people who now have to act. Put the changed fact on that object and everything connected to it lights up. Nobody had to ask which open deals this regulation touches. The answer was already drawn on the same map.

Correctness comes first, and it is proven

Before an agent can see anything it has to be standing in the right place. It has to act on the right company, the right person in their current role, the latest accepted fact. It sounds like a low bar. It is where most deployments fall over.

At a deep-tech company the partner list was a dirty spreadsheet of about 2,700 companies. We rebuilt the territory from government sources, hospital lists and integrator registries, linked it to their CRM, and the list collapsed to 64 ranked accounts. In a climate-reporting build we started from 4,437 companies in a public CO2 register and found company names too unreliable to dedupe on; the domain became the key. At a planning-software company a new head of sales changed the deal stages without telling the CFO. An agent working from that CRM that week would have been working from a fiction.

Correctness is the meeting-prep agent that runs at 9am and pulls the previous calls, the emails and the deal history for the right account, so a rep checks it before every external meeting. That works, and it is in use daily. It is also table stakes.

Sight is what connection adds

At the planning-software company we built a list around an end-of-life open-source planner that a lot of companies still run. On GitHub you can see who commits to it and who forks it. Stars mean little; people bookmark things. But when several people from the same company commit, it is a reasonably safe bet the company runs it in production. Job boards show whether that company is still hiring for the skill. The CRM shows whether anyone has ever spoken to them.

Each source alone says almost nothing. GitHub says someone at this company writes code against the planner. The job board says they want a scheduler. The CRM says no record. Put the three on one company object and they say something none of them contains: this company is running a dying planner in production, is still investing in it and has never heard from us. That is a sentence no source wrote. The connection wrote it.

The same thing happens with a competitor campaign. We send an instant Slack ping when a competitor launches a new ad campaign. On its own that is a marketing curiosity. On the graph it lands next to the accounts where that competitor is already in the deal. Now it says: these deals are about to hear a new pitch.

This is why funding rounds and job postings alone stopped working as signals. Everyone gets them and inboxes cannot cope. A signal is only worth something once you know what it is connected to. The agent is signal-agnostic. The graph decides what the signal touches.

Search cannot give you sight, because you have to know the question

Retrieval returns the documents you asked for. Sight is the connection you did not know to ask for. Nobody on that sales team would have typed “companies where several employees commit to this repo, that also posted a scheduling job, that are not in the CRM”. The question does not exist until the sources are side by side. Ortelius could have asked whether the coasts fit. He did not, and nobody else did, until the coasts were on one page.

A search index is a good librarian. It brings you the book you name. It does not notice that two books on different shelves describe the same coast.

The obvious reply is that a smarter model will just read everything and notice. It will not, because the context window is temporary working memory. A model that reads the GitHub activity and the job postings this morning has no memory of them this afternoon, and cannot tell you later which of its conclusions were checked. The connection has to live somewhere, attached to the things it connects, with its sources. That place is the world model, and the language model can be rented on top of it.

Sight is only worth having if you can check it

Ortelius credited the people whose maps he used. That was not politeness. It meant anyone could go back to the source and see whether the coast really looked like that.

We build to the same rule. Every accepted fact on the graph carries provenance: where it came from, when it was true, when it was recorded. When a new fact replaces an old one it supersedes it. Nothing is overwritten in silence. And what an agent concludes stays separate from what the company has accepted as true. The line “this company is migrating off the planner” is an agent’s interpretation. It sits on the company object, marked as such, above the commits from several people, the job posting with its date and the empty CRM record it was built from. A rep can open it and check. If the conclusion were just written into a CRM field, nobody could tell later whether it was seen or guessed.

A connection you cannot trace is a rumour with a confident tone. A connection with its sources attached is something you can act on, and something a second person can disagree with on the evidence.

The whole thing is built for sight: connected facts that show what no single source contains.