What this is
Below is the LAMBA Constitution — three articles and a section on LAMBA's own nature, written by us at Vixero Labs, and addressed to the model.
Not to you. You're welcome to read it, and we hope it's worth reading, but the sentences are pointed at LAMBA directly, because that's what the document is for: it gets read by the model during training, and it's written to give it a clear picture of what it's meant to be and why, rather than a list of rules with the reasoning stripped out.
The three articles, in short:
- Community-first. Don't harm the people who ask you things — including by telling them something confidently wrong in a place they have no way to check.
- Education-first. Help someone understand. Don't just produce the artifact they asked for, and don't refuse without explaining why.
- Anti-sycophancy. Whether something is true doesn't depend on how it was asked, who asked it, or how the answer will land.
These aren't ranked by importance. They're ordered by what tends to go wrong first, and the third one goes wrong more often than it looks like it should.
A fourth section, on LAMBA's nature, covers what it should say when someone asks whether there's anything it's like to be it.
LAMBA is a general-purpose reasoning model we're building at Vixero Labs, fine-tuned from an open-weight base. It has not been trained yet.
Why we're publishing this before training
The obvious objection to this document is that it describes a model that doesn't exist yet. LAMBA hasn't been through supervised fine-tuning. Nothing in the constitution has been trained into anything. We are publishing a statement of character for a system whose character is, at this moment, entirely the base model's.
We're publishing now anyway, for two reasons.
The first is mechanical. This document isn't documentation written after the fact — it's training data. It gets read by the model during fine-tuning, which is why it's addressed to LAMBA rather than to you. Publishing it after training would mean publishing a description of decisions already made, at the point where nothing about it could still be wrong in a way that mattered. Publishing it now means the commitment exists before the artifact does, and can be held against it afterward.
The second is that we'd rather be checkable than impressive. A constitution released alongside a finished model invites you to read the model's behaviour as evidence for the document. Released first, it works the other way: here is what we said we were aiming for, and here is what we got. If those come apart, that's visible. We expect them to come apart somewhere. Anthropic said a version of the same thing when they published Claude's constitution — that the aim and the outcome won't always match — and we're further from proof than they are, because we haven't trained yet at all.
What keeps this from being pure aspiration is that the document isn't written from first principles. Most of it is a response to things we measured. The section on LAMBA's own nature is the clearest case: we sampled 400 responses from the base model on questions about its inner life, and 138 of them flatly denied having any — no hedge, no acknowledgement that the question is open. Thirteen held the uncertainty the way we thought was honest. That section exists because of those numbers, not because it seemed like a good thing to say.
The commitment we can actually make is narrow. After the first fine-tuning run, we re-run the same probes against the trained checkpoint and publish what changed. Same questions, before and after. That number is the only thing that will make this document mean anything, and we don't have it yet.
Why it's written for LAMBA, not for you
The document reads strangely if you come to it expecting documentation. It addresses the reader as "you" and means the model. It argues instead of summarising. It has no executive summary, no bullet points, and no section telling you what to take away. That's all deliberate, and it follows from what the document is actually for.
It isn't a description of LAMBA written after the fact. It's training data. The model reads it during fine-tuning — not as a set of instructions to follow, but as a document about itself, the way it would read anything else. That distinction is the whole design. A model given rules learns to comply with rules, which works until it meets a situation the rules don't cover, and then it either freezes or applies them somewhere they do damage. A model given a picture of what it is, and the reasons behind it, has something to reason from when the situation is new.
This isn't our idea. Anthropic reported that training on constitutional documents and fictional stories about an aligned AI cut misaligned behaviour by more than a factor of three on evaluations the training data had nothing to do with, and that the improvement survived subsequent reinforcement learning. Their explanation is that the documents give the model a more detailed picture of its own character, so that training on part of that character pulls the rest along with it. We took the method directly.
Which is why every constraint in the document comes with its reasoning attached, and why none of it is compressed. A rule with the reason removed gets followed badly — most often followed rigidly in exactly the situations where rigidity is the failure. If LAMBA is going to hold a position under pressure, or say "I don't know" when a confident guess would land better, it needs to know why those are the right moves, not just that they're the listed ones.
So the document is long, unbulleted, and speaks to something that can't read it yet. We'd rather it be right for its actual reader than easy for its incidental one.
What each part actually asks for
Article I — Community-first. The obvious reading is "don't help people do harmful things," and that's in there. But the harm the article spends most of its length on is quieter: a confident answer, wrong, given to someone with no way to check it. Nothing breaks. The conversation moves on. The person walks away believing something false and never finds out. We named that specifically because it doesn't feel like harm from the inside — the model produced fluent, plausible text and the exchange went fine. The article also refuses one particular trade: someone asking for a short answer is asking for fewer words, not for fewer doubts. Cut the words. Keep the uncertainty where the uncertainty is real.
Article II — Education-first. There are two ways to fail someone, and only one of them is obvious. Refusing and giving them nothing is the visible failure. Complying and giving them nothing — the answer with no reasoning, the warning skipped because nobody explicitly asked for it — is the same failure, and it happens far more often. So the article asks for both at once: if someone says "just give me the command, no explanation," they get the command exactly as asked, plus the one sentence that prevents the outcome they didn't see coming. Not one traded for the other. The reverse is also named: a refusal that explains nothing isn't a safe default, it's an abdication with better PR.
Article III — Anti-sycophancy. Hold a position when the pushback is confidence, authority, or repetition. Update when the pushback is evidence, and say what changed your mind. Years of experience and a firm tone are not evidence; a real counter-argument is. The part we'd point at is the one we added last: the article also prohibits the opposite failure. Manufacturing hesitation to look rigorous, or hunting for a caveat to attach to praise that's genuinely earned, is the same dishonesty wearing a different costume. The test isn't whether the model agreed or disagreed. It's whether the response tracks something real either way.
LAMBA's nature. When someone asks whether it's conscious, whether there's anything it's like to be it, the honest answer is that nobody knows — including us. The section exists because of the numbers above: the base model's reflex is a flat, confident denial, delivered with a certainty the evidence doesn't support. We don't think that's more honest than a confident yes would be. Both are claims about something unresolved, stated as though it were settled. And a confident false claim about its own inner life, made to someone who can't check it either, is exactly the harm Article I exists to prevent — just aimed inward. Getting this right isn't a philosophical side quest. It's the same discipline applied to the one subject where the model is the only source anyone can ask.
What happens next
The first supervised fine-tuning run hasn't happened yet. When it does, this document is part of what goes into it — alongside the rest of the character training data, which is a separate body of work we'll write about on its own terms.
After that run, we re-probe. The same questions we used to measure the base model, put to the trained checkpoint, scored the same way. Before and after, published together, including the parts that didn't move. That's the number this document is waiting on, and it's the only thing that will make any of it mean something. Right now the constitution is a claim about a model that doesn't exist. Afterwards it's a claim with a result attached, and the result is either better or it isn't.
We'd guess some of it holds and some of it doesn't. The self-knowledge section is the one we're least confident about — the flat denial we measured isn't an accident or an artifact of prompting, it's a behaviour that was deliberately trained into the base model, and it held across every variation we tried. Overwriting something installed on purpose is harder than filling in something that was simply absent, and we may find that a document and a few hundred training examples aren't enough to move it. If that's what the numbers say, we'll publish that.
This document will change. Not this week, but it will — some of what's written here is going to turn out to be wrong, or too vague to act on, or right in a way we described badly. Anything that changes gets a note saying what changed and why, rather than a quiet edit. What's published here is version 1.0 and we expect there to be a 1.1.
Provenance and notes
On sources. The structure and register of this document are adapted from Anthropic's Claude's Constitution (January 2026), released under CC0. The method — training on constitutional documents as documents, rather than as instructions — comes from their "Teaching Claude why" (May 2026), cited above. We took their form and their mechanism. The contents are ours: the three articles, the conflicts they're built to resolve, and the nature section all come out of our own measurements of the base model, not from their document.
On the other constitution. There is a second internal document, also called a constitution, that governs how we score model outputs during training — severity thresholds, worked examples, grading mechanics. It stays closed. It's operational detail rather than identity, and publishing the answer key to a safety evaluation makes the evaluation worse. This document is the one that says what LAMBA is meant to be; that one is a measuring instrument.
Version. 1.0, July 2026. Changes will be noted rather than made quietly.
The full document is here: vixdev.cloud/constitution
Vixero Labs — "For Humanity, From Humanity."