The Machine-Readable Estate

01 A portfolio of operating companies is worth the sum of its parts. A portfolio whose companies, people, capabilities and history resolve into one machine-readable graph is worth more than the sum, because every addition attaches to everything already there.

02 Lines of code is the wrong measure and an unreliable benchmark. There are no audited disclosures to compare against, and repository boundaries, generated code and vendored dependencies make cross-company comparison meaningless.

03 The right measures are countable: canonical entities, relationships per entity, the share carrying checkable provenance, the number of independent sources, freshness, and how much of it a machine can read without an account.

04 The scarce input is not content. It is entity resolution — the discipline that the person in an acquired archive and the founder in the portfolio are the same identifier, asserted by somebody, on a date.

05 The firm instruments its own estate against these measures and publishes the result, including where it is failing. The first reading found thirty records that share a file format and are not yet one graph.

Ask how large a technology estate is and the answer comes back in lines of code. It is the wrong instrument, and not marginally. There are no standardised, audited disclosures of code volume for any company of consequence, so every comparison is between one number somebody counted carefully and another they did not. Repository boundaries move code between the numerator and the denominator without anything changing. Generated clients, vendored dependencies, test fixtures and twenty years of history all count or do not count depending on who is asking.

The deeper objection is that the measure does not track the thing that produces returns. An estate of a hundred million lines that is mostly applications, each with its own data model and its own idea of what a customer is, is a hundred integration projects nobody has funded. A far smaller estate whose records resolve into one connected graph can answer questions that the larger one cannot answer at any price.

This is not a technology argument. It is the same observation that made standardised financial reporting valuable: the accounts of two subsidiaries are worth more when they are prepared on the same basis, and the value comes from the comparability rather than from the volume of the ledgers.

A holding company accumulates the usual assets: operating businesses, domains, archives, directories, conference histories, research libraries, customer and counterparty records, transaction history. Most groups hold these as separate estates with separate systems, and reconcile them, badly, when a transaction requires it.

The alternative is to hold them as one graph from the beginning. A speaker at a conference the group owns is the same entity as a founder in the portfolio, who runs a company that appears in an industry directory the group acquired, which uses infrastructure the group operates. Each of those is an edge, each edge is asserted by somebody on a date, and the resulting structure is queryable.

The distinction is not cosmetic and it shows up in valuation. Two hundred websites is a portfolio, and a portfolio is worth what its parts fetch. One graph with two hundred public front doors is an institution, and it gets more valuable every time something is added — because a new entity can attach to everything already present, and a new website cannot attach to anything.

This is also the honest answer to why the work is hard. Publishing records is easy and every group could do it this quarter. Resolving entities across records written by different people, in different decades, about the same person is the part with no natural deadline, because nothing breaks if it is skipped.

If the graph is the asset, an acquisition is not valued on its revenue multiple alone, and certainly not on its code. A twenty-year-old industry publication is thousands of articles naming companies, people, financings, events and relationships, accumulated by journalists who verified them at the time. That is an entity history with provenance attached, and it cannot be recreated by any amount of present-day effort, because the events are over.

What is bought is information density, authority, entity history, audience and unique data. What is inherited is a content management system nobody wants. Groups that buy publications for their traffic tend to discover the traffic was the least durable thing in the transaction.

The discipline is to run every acquisition through the same sequence and to publish which step it reached. Acquire, preserve, structure, entity-resolve, attach provenance, connect to what is already held, expose at a machine-readable address, and finally make it answerable by an agent without a human in the loop.

Steps four and six are the whole exercise and the two that get skipped, because the site still renders without them. An organisation that skips them repeatedly ends up with a content operation and a conviction that it has built a knowledge base.

Six numbers describe a knowledge estate, and they are deliberately not combined into an index. A single score lets one dimension collapse while the average holds, which is precisely the failure the score was bought to detect.

Canonical entities is how many distinct things are named. Relationships per entity is how connected they are. Provenance is the share of assertions naming an asserter that the graph itself defines — not merely a filled-in field, because a citation to an identifier nobody has declared is a citation to nothing. Independent sources is how many separate records contribute. Freshness is the share of decaying claims still current. And machine accessibility is how much can be read without an account.

The sixth is the one that separates a connected estate from a collection of tidy ones: the share of relationships that cross a property boundary — an assertion made by one record about an entity a different record defines. Every other measure can be improved by one property talking more about itself. Only this one improves when the estate becomes more connected.

The firm applies these measures to its own estate and publishes the result rather than describing the method. The first full reading, taken across thirty-six repositories on the branches they actually deploy from, found 1,422 canonical entities and 2,726 declared relationships contributed by thirty records.

It also found that 8 per cent of relationships cross a property boundary, and that no entity at all is declared by more than one property. On the measure that matters, the estate is thirty records that share a file format and are not yet one graph. Provenance stands at 9 per cent, for a single repeated reason: the automated agents that sign the records are identifiers no record declares, so thousands of otherwise sound assertions cite an entity that formally does not exist.

Both findings are cheap to fix and neither was visible before something counted. That is the argument for instrumenting this at all. An estate that believes it is connected, and has never measured whether it is, will continue believing it for as long as nobody asks — and the belief costs nothing until a transaction depends on it.

The reason to publish an unflattering first reading is the same reason to publish an audited account rather than a projection. A number a reader can check is worth more than a number they are asked to accept, and an organisation that only publishes numbers when they flatter has told the reader how to treat the flattering ones.

Three signals indicate a group is building this rather than describing it. Records published at machine-readable addresses without a login, because a knowledge estate behind an account is a product feature rather than an asset. Identifiers that are stable across properties, so the same person is the same identifier in the conference archive and the portfolio register. And a published measurement, taken on a stated basis, that has moved.

The counter-indicators are equally clear. A data lake with no shared identifiers. An artificial-intelligence initiative announced before any record was made legible to it. And any claim to be the most machine-readable organisation in a category, which is unverifiable by construction and therefore tells the reader only that nobody checked.