Anyone Can Copy the Methodology. Only One Lab Can Keep the Promise.

Vinci ships August 8 - one variant, vinci-studio, built on a strong open base through our Constitutional Fine-Tuning process, shipped open under Apache 2.0 with a public Constitution, Character document, methodology paper, and adversarial test results run against the deployed weights.

The SimpleDirect blog has the company version of what that is.

This is the version I'd write for another founder, about why I'm giving away the part everyone assumes is the secret.

Here's the whole argument in two sentences.

Anyone can copy the methodology - and they should, because I'm publishing it.

The thing they can't copy is whether the model actually does what the constitution says it does, in public, under tests anyone can run.

That second sentence is the moat.

Everything else in this essay is me working out why.

The thing I kept noticing

Open Hugging Face today and you'll find hundreds of fine-tuned models.

Some built on Llama, some on Qwen, some on the same DeepSeek base we use.

Most do similar things. A handful do specific things well.

The competition between them isn't capability.

There's too much capability, sitting right there, free.

The competition is trust - whether you believe what any of these models do enough to put it in front of your customers.

I learned this in rooms with regulated buyers.

They weren't reading benchmark scores to me.

They were asking: how does it behave when it's wrong? Where does it break? What happens when my compliance team has to audit a decision it made?

Capability mattered, but only as a threshold.

Below it, no model was acceptable.

Above it, the deciding factor was never the parameters.

It was whether the thing in front of them did what it was claimed to do - and whether they could check.

That's not a model problem.

That's a fidelity problem.

And fidelity is a different game than the one most AI companies think they're playing.

Why I'm giving the recipe away

For a while I thought the methodology was the asset.

The Constitution took six months. The Character document took weeks.

The process for turning a strong open base into something well-behaved and consistent is genuinely hard-won.

So the instinct is to guard it.

I'm doing the opposite.

The methodology paper goes out in full on August 8, and I want competitors to read it.

Because a methodology is a recipe, and recipes copy in an afternoon.

A recipe lists the steps.

It cannot make you actually cook the dish the same way every single night - when it's expensive, when the customer is angry, when honoring it costs you a deal you wanted.

The hard part of cooking was never the recipe.

It's the line, every service, under pressure.

Values work the same way.

They're not a moat because they're mine - every company can write "we value safety and honesty" into a document and most do.

That's a poster, not a position.

A competitor copies the words and now we both claim values, and nothing is defended.

Values become a moat only when fulfilling them is a behavior you have to perform - repeatedly, under cost, and get caught if you fail.

The constitution is a promise.

Keeping it is the work.

And the gap between the promise and the kept promise is exactly what I'm making visible.

That's what the verification bundle is for.

It turns values from a claim into a test you can fail in public.

I publish the constitution that says how the model should behave.

I publish the adversarial results showing how the deployed weights actually behave.

Anyone can line them up.

If they don't match, I don't get to talk my way out of it - the bundle is the contract, and the contract is public.

Giving the recipe away isn't a contradiction with the moat.

It's the precondition for it.

By publishing the method, I remove "we have a secret sauce" as an excuse and force the whole competition onto the one axis nobody can copy: did the model actually turn out the way I said it would.

Why this is the right bet for three people

I can't compete on parameters.

I'm not training a frontier model from scratch - I don't have the compute, the team, or the reason to.

Strong open bases exist; I build on them.

But once everyone is building on the same open bases, the base stops being where you win.

We'll all be standing on similar ground.

What separates the companies on that ground is whether their model does what their documentation promised - and whether they'll stake something on it.

That's a fight a three-person team can win, because it isn't won with money.

It's won with discipline you can verify.

A well-funded competitor can copy my methodology, match my benchmark scores, and outspend me on everything - and still lose the regulated buyer, if their deployed model drifts from their own spec and mine doesn't, and the buyer can see the difference.

Map it to the layers I actually think are defensible.

The model is shaped to our constitution and character, not a generic one - that's identity, and a copied recipe doesn't reproduce it.

The 30-day fix promise means I pay, in public, when fidelity breaks - that's stakes, and a competitor can copy the sentence but not the bill.

The bundle is the contract and I'm on the hook for the gap - that's accountability, and it's measurable.

None of those is the methodology.

All of them are downstream of actually keeping the promise the methodology lets me make.

The part I have to be honest about

This moat is exactly as strong as the verification is hard to fake.

No stronger.

If the adversarial tests are weak, or I'm grading my own homework, "demonstrated fidelity" quietly collapses back into a values poster - and I'd deserve to get caught.

The whole position rests on the tests being genuinely adversarial, runnable by third parties, and run against the weights I actually deploy rather than a cleaned-up demo.

So I'm staking it there, out loud.

The bundle isn't a marketing asset that happens to look rigorous.

Its only job is to be hard enough that passing it means something.

If the tests are soft, the moat is fake, and a sharp reader should poke exactly that hole.

I'd rather name it than have it found.

This is also where I think the closed-weight labs and I actually diverge, and it's worth being precise about it.

OpenAI, Anthropic, and Google clearly understand that the model isn't the differentiator - their specs, safety work, and documentation are the real assets.

But they keep that work closed and ask you to trust the result.

Their bet is asked-for trust, backed by brand and scale.

Mine is the opposite.

Publish the method, expose the deployed behavior, and let verifiable fidelity do what their brand does for them.

It's a different bet, available to a small company precisely because it doesn't require their resources - only their discipline, made checkable.

Whether it beats asked-for trust is the open question.

I think it does, for the buyers who've been burned by taking behavior on faith.

What I'm watching for

By October I want one thing answered, and it isn't a benchmark.

Does a regulated buyer who reads the constitution, the character doc, and the adversarial results actually trust vinci-studio enough to deploy it - and does that trust come faster than it would from a closed provider asking them to take it on faith?

Two sub-questions under it.

Does the 30-day fix promise hold the first time we hit a real divergence - because we will, and how I handle it is the entire brand.

And are the tests hard enough that passing them moves anyone, or do I find out my verification was theater the moment someone serious runs their own.

If the answers are yes, the position holds - and it's a position a well-funded competitor can't simply buy their way past.

If they're no, I learn something expensive and update in public, which is the only way I know how to do this.

vinci-studio ships August 8 with the whole bundle.

Read the constitution, run the tests against the open weights, see if it does what I said it does.

That's the bet.

Not that the method is secret - that the promise is kept, and you can check.