The governed intelligence layer
for real businesses.

We build company AI, the infrastructure beneath it, and the intelligence that accumulates when real businesses do real work on it. That is one company with one motion, not a holding structure with three logos.

The public shape is simple and it is deliberate. Sabina is what a customer hires. Forge is the governed platform and operating spine she runs on. Hyperion is a specialized standalone trades product hosted under Forge. The company is Dominus Foundry, family-held, built and run from Albuquerque.

Company AI becomes infrastructure. Infrastructure becomes intelligence. The rest of this page is the argument for that sentence, with what is real separated from what is not.

Employees can teach her. Leadership decides her authority. The company can change that authority or take it back. She is hired into a business, taught how that business works, and given authority in writing — and she becomes increasingly specific to that company through governed memory, teaching, correction, workflows and business context.

She holds a seat the way a person does: her own address, her own line, her own scope of work. She is brought on and onboarded rather than installed, because the thing that makes her useful is not the software — it is what her employer has taught her and what they have authorized her to do about it.

She becomes company-specific through governed memory and teaching. That is not a claim that every customer receives a separately fine-tuned foundation model.

Capability stopped being the constraint some time ago. What stops a business putting AI on its own phone line is not whether the model can hold the conversation — it is that nobody can say in writing what it is permitted to do, or show afterwards what it did and under whose authority.

Consequential actions route through authority granted by the company, in writing. Authority can differ by employee, role, customer, object, condition and responsibility. The authority model is durable and auditable, and withdrawing a grant stops future work under it.

Revocation withdraws permission for future work. It does not unsend a message or reverse a completed action, and this page makes no claim that it does. The execution path described here is in acceptance testing — see what is already real.

Two companies in the same trade, on the same street, do not run the same way. Their escalation rules differ, their customers differ, their tolerance for a judgment call differs. Generic competence cannot close that gap, and prompt engineering closes it only until someone changes the prompt.

So the useful unit is not the model. It is what one company has taught, corrected and authorized — accumulated over months, attached to outcomes, and owned by them. That is also the part a competitor cannot copy by shipping the same feature, and the part a customer cannot take to a rival product by exporting a table.

An AI employee needs somewhere to work. Forge is the governed technology platform and operating spine: the application infrastructure, the data and workflow layer, and the system of record she acts on rather than visiting through a narrow integration.

It is also why the proof is real rather than a pilot. Forge has been in production at a commercial roofer since March 2026, and is also used inside a founder-affiliated commercial AV and security integrator. Sabina runs on that foundation.

Forge is infrastructure and specialized tooling. It is not the company-level thesis, and this pitch is deliberately not a catalog of its modules.

This is the one place on this page where status is exhaustively qualified. Every other section reads from it, and nothing elsewhere states a stronger version of any row.

Talk to the customer, not the pitch. All Weather Roofing — on Forge since March 2026, invoiced and paid. A reference call with the owner of All Weather Roofing is available on request. Ask for the reference call →

The structural precedent is not field-service software. It is data and intelligence infrastructure — companies whose durable value came from a standardized proprietary record created inside the work their customers were already paying for, rather than from the workflow product itself.

Verisk is the clearest version of the pattern: data contributed by customers and data generated inside the transactions its products support, turned into risk analytics embedded back in the workflow. CoStar is the other half of the argument — the enduring value of a comprehensive standardized record, assembled from research, public, user and acquired sources. Our thesis applies that pattern to business decisions rather than claims or properties: decisions joined to authority, evidence, alternatives and outcomes.

These are business-model analogies. They are not evidence that our future intelligence products or their margins already exist — the network layer is described in the pending application and is not operating today.

1. Residency

   She works on the system of record rather than visiting one task through a narrow integration.

2. A mandatory governance path

   Consequential effects traverse the authority and execution boundary rather than going around it.

3. Decision chronology

   The record is created before the outcome, which is why it cannot be manufactured afterwards.

4. Company-specific institutional learning

   Months of teaching, correction and accumulated operating context, owned by the employer — and the switching cost that useful company memory creates.

5. Owned inference path

   Open weights on our own hardware, with no third-party inference provider in any customer path. We control the serving and training path needed to turn governed episodes into durable improvement, and the cost curve that comes with owning it.

6. Cross-company pattern discovery, where permitted

   Where contracts, permissions and governance permit it, eligible derived patterns across companies can support discovery no single company could reach alone — returned to each as its own intelligence. Described in the pending application; not operating today.

7. Filings and real deployment

   A pending application over the mechanism, and a substrate already carrying paid production work.

Point agents know the conversation. Vertical platforms know the transaction. What neither preserves is the governed decision in between — and the company that has been accumulating it since the first instrumented customer is the one holding the asset.

19 provisional applications filed with the USPTO, filed March–August 2026, covering the governed execution, authority and intelligence architecture. Patent pending: filed, not issued.

U.S. non-provisional application 19/791,511 was filed with the USPTO on 30 August 2026, naming Mark F. Lord as inventor. Patent pending: filed, not issued. It describes proof-carrying governed execution of autonomous work, and the contamination-resistant derivation of decision-grade operational intelligence across a multi-company panel — the mechanism behind both the governance argument and the compounding one.

⛔ Stated precisely, because the difference matters in diligence: these are filings, not granted rights. Nothing here is patented, issued or protected by an issued patent, and nothing on this page claims a competitor is blocked. The correct reading is patent pending.

A limited number of Founding Seats, awarded by application. A company completes a readiness assessment; we review it; there is a call where the employer writes the job description — what she will handle, what she must never do, who supervises her, how escalation works. Then they interview her, provisioned with their own material, and she asks qualifying questions back. If there is a fit, they bring her on. Founding Employers help polish the product through real use.

Pricing is custom to each company, built from what the interview establishes about the job, and sent afterwards as a written proposal. There is no rate card and no published figure on any surface — which is also why the qualification work sits in the readiness assessment rather than in a price a visitor self-selects against.

The comparison against a human hire is the one number a buyer already knows. A fully loaded front-office coordinator — wage, employer tax, workers’ compensation, health, retirement, the seat — runs about $5,500 a month, roughly $32.50 an hour at the desk, converging with the BLS Employer Costs for Employee Compensation figure of $5,474 a month.

That is the modeled cost of a human hire and an anchor assumption, not an observed customer figure — and not Sabina’s price, which is quoted per company and appears on no surface.

The conservative case assumes no raise and one new company a month — the throughput two founders can actually onboard — and it models the same 2% monthly logo churn the funded case does, roughly a fifth of the book a year (about 22%). It reaches 25 seats and $971K of recurring revenue at month 36.

The point of publishing it is what it does not do: with churn in the model it still never runs out of money. The trough is in month 5, it is cash-flow positive from month 6, and it ends month 36 with roughly twenty-seven times the cash it started with. That is the whole claim — not that this is the plan, but that a no does not end the company.

Modeled, not observed — Dominus_Foundry_ProForma.xlsx base scenario, re-modelled at 2% monthly logo churn per Mark 2026-09-16 · derivation: scripts/floor-churn-model.mjs (zero-churn control reproduces the workbook, 252/252 cells), as of 2026-09-16. It is the floor, not the plan. It assigns $0 of revenue to the future network business.

The funded case holds revenue per seat constant and the churn assumption identical, and changes one thing: how many companies we can bring on. It reaches 94 seats and $3.49M at month 36.

Modeled, not observed — Funded scenario modeled 8 September 2026; not yet in the 30 August workbook. Neither scenario is a new pricing quote, and the funded case also assigns $0 to the network business.

Models crossed the line where they can carry consequential business work about eighteen months ago. What did not arrive with them was any way for an owner to say, in writing and in advance, what the system may do — and to show afterwards what it did and on whose authority. That gap is why so much corporate AI is still confined to drafting.

Meanwhile employees are already using unsanctioned AI on company work, because the sanctioned internal option is weak or does not exist. Every month that continues, a business accumulates exposure and no record. The companies that will own the operating record of the next decade are the ones instrumenting it now, before the volume exists to make it valuable.

The other half is cost. Serving self-hosted open weights for a book of companies is affordable at a scale it was not eighteen months ago, which is what puts the inference layer beneath the product on our own hardware rather than a third party’s — and it is a line item in this round.

Founder-led today. The round adds the first account executive and deployment lead, so selling and onboarding become systems rather than founder-only work.

Family-held, and built and held by the Lord family. Fide et Familia. That is the company’s identity and the reason it is built for a long hold — it is not the investment thesis, which is everything above it.