Own permanently. Give ownership up only for something worth more than it.
The charter of this office declares six functions, and ownership is one of them. This is what that function does: it decides how the estate enters a position, and it decides the one question that actually determines what this office is worth in thirty years — when to stop owning something.
Most groups answer that question by default. A fund answers it with a date. A founder answers it with an offer. This office answers it with a test, written down before the offer arrives, because a test written after one is a rationalisation.
There are six ways this office can take a position, and they are not interchangeable. Each has a condition that must hold before it is the right one, and the failure mode of a holding company is reaching for whichever is most available rather than whichever is correct.
Build — when a vertical deserves a company.
Not when a product deserves a feature, and not when the estate has a gap on a slide. A vertical deserves a company when it has customers who are not us, economics that stand alone, and someone who will run it as their own thing. Everything short of that is a programme, and a programme inside a holding company is a cost centre with a founder's title on it.
Acquire — when the assets strengthen the mesh.
The test is not whether the business is good. It is whether owning it makes something the estate already owns work better — a distribution surface the mesh can settle across, a dataset an agent can read, a licence that removes a dependency. A business acquired for its own returns is a fine investment and belongs in the ventures arm. A business acquired because it makes six other things stronger belongs here.
Standardise — when interoperability raises the value of what is already owned.
A standard is given away, always, and the reason is commercial rather than generous. An organisation deciding whether to build against a format is underwriting a dependency, and the risk it prices is that the owner changes it, charges for it, or withdraws it. Removing that risk is worth more in adoption than control is worth in optionality — so the specification is open, the licence is permissive, and the position is the best implementation of it rather than ownership of it.
The test is narrow: does a second organisation adopting this make an asset this office already holds more valuable? If the answer is no, it is a side project with a licence file.
Publish — when a public good raises discovery.
Some things are worth building and worth giving away outright: a definition, a register, a tool that grades anyone's site including ours. They earn their cost in being found and cited rather than in revenue, and they are held to a stricter standard than the commercial properties — a public good that flatters its author is an advertisement, and is worth nothing as either.
Partner — when someone else has what cannot be efficiently reproduced.
Distribution into a market this office does not reach, physical infrastructure, a regulated licence, an expertise that took twenty years. The test is the word efficiently: almost anything can be reproduced given enough time and capital, and the question is whether reproducing it is a better use of both than the terms of the partnership. Where it is not, the partnership is correct and the instinct to own everything is the expensive one.
Sell equity — only when what is gained is worth more than the ownership surrendered.
This is the only one of the six that is subtractive, and it is the one most often decided by circumstance rather than judgment. Capital, a partner, or distribution can each be worth more than the percentage they cost. Frequently none of them is, and the round happens anyway because it was available and the market was warm.
The comparison is against holding, not against zero. A position sold at a good price is still a position that has stopped compounding, and the buyer is generally paying for the part that has not happened yet.
The reason to hold rather than sell is not sentiment about companies. It is that a company retained keeps compounding inside the system, and a company exited leaves it — taking with it the thing this office is actually assembling, which is not a portfolio but an estate that shares one set of rails.
A controlled company can inherit a capability by default. A minority position cannot be told to adopt anything. That is the whole argument for owning most of a smaller number of companies rather than a little of many: inheritance only works downhill, and the fourth move of the thesis depends on it entirely.
The category has an institutional definition, and it is not written here: a machine-native holding company is defined by the merchant bank, and the mechanism it turns on is the propagation premium. One claim, one canonical home — the definition belongs there, the evidence belongs here.
This office is being built as a machine-native holding company. It is not one yet, and the difference is checkable rather than a matter of opinion. The definition carries seven conditions, written so that a third party can test any holding company against them by fetching public URLs, without that company's cooperation. Below is where this one currently stands, including the conditions it fails.
No projection of what any of this becomes is published here, and that is deliberate. This office holds that there are no invented numbers, at any scale, and a look-through valuation built on companies that do not yet exist is the largest invented number available to it. What is published is the register, with the state each property is actually in.
Assessed 30 August 2026. A condition is met only where an outsider could confirm it by fetching a public URL — not where the intention exists, and not where it is true of some properties in the estate but not of the parent.
1. Resolvable identity for the parent and each controlled company. Met.
The parent now serves a permanent identity URI at /organizations/gord-holdings, resolving to a record with its asserter, its date and its expiry. It was the last property in the estate to do so, having been the most conspicuous failure on this list — the entity a machine resolves first, answering nothing.
Serving it surfaced a question the register had never been asked. Three properties render from one repository and each announced an organisation named after its own domain, while the fragment defined one — so two organisations were announced on the mesh and asserted by nobody. Nothing surfaced that for as long as an identifier only had to exist; it became visible the moment one had to resolve to a record. Settled: the capital and ventures arms are brands of this office, not organisations. All three properties now announce the one that exists.
2. Ownership published as structured data. Not met.
The register names every property and the state it is in, and publishes no ownership or control edge at all. What is published is what this office operates, which is a statement about itself. The structure is the fact this entity exists to record, and it is currently prose.
The mechanism now exists and the terms do not. controls was added to the profile for this case: it states which organisation answers for another with no percentage disclosed, so a structure is publishable by a group that will never publish a cap table. It requires a basis and a date, both of which live in constitutional documents rather than in this repository, and inventing either would put a fabricated structural claim into the one graph whose value is that it can be trusted about exactly this. So the fragment publishes the four organisations waiting on those two facts, rather than leaving the gap as silence.
3. A machine-readable statement of what each organisation does, and who answers for it. Met.
Every organisation in the estate publishes a charter declaring its functions, the threshold above which a human must approve, and the named person accountable. This one included, and it was the last to conform — which was the first finding of its own audit.
4. An agent-readable interface whose operator resolves. Not met.
No property in the estate publishes an agent card. Nothing here would be improved by publishing one for an agent that does not run: a card naming an operator for a service nobody can call is the unverifiable claim this group audits other people for.
5. A shared identity and coordination layer, verifiably shared. Partly met.
Every property publishes the same handshake, the same charter format and the same front door, and any of that can be checked from outside. But the shared code is copied into each repository rather than installed from one place, so an improvement made once does not arrive anywhere else — and an audit of that this week found three copies of one validator with the estate's own flagship claim type missing from three of them. Adoption of the formats is real; inheritance of the capability is not yet.
6. One canonical home per fact, dated, attributed, expiring. Partly met.
Every assertion this estate publishes carries an asserter, a date and an expiry, and that has been true for some time. What is new and not yet in use is the channel that makes a fact change reach the properties that render it; the format exists and no property holds a lock against it yet.
7. Agents operating, with a public record a stranger can recompute. Partly met.
Agents run here and the work is recorded in public. The record is not yet independently recomputable end to end, which is the part that matters: a log a reader has to take on trust is a log, and the condition asks for evidence.
Two of seven met, three partly, two not. That is the honest state of a category this office intends to occupy and does not yet occupy, published at the number it is rather than the number it would like. The conditions were written to be failed by whoever wrote them first; anything else would be a positioning claim with a checklist attached.
Ownership is not the reward for building something. It is the mechanism by which the next thing is cheaper to build.