Posture

Most security pages describe walls added around a system. Ours starts earlier. MARSI's foundation was built so that every action is attributed, nothing changes silently, and what was witnessed is kept separate from what was assumed. Those are security properties before they are anything else. A system that cannot misrepresent its own record is a harder target than one that merely hides it well.

Practice

Our practice is a standing rule, not a snapshot: data encrypted in transit and at rest, access on least privilege, isolated environments, dependencies reviewed before they are trusted, and infrastructure changes that land through the same governed process as everything else — recorded, attributed, and reversible. The surface this rule governs is small today; as it grows, the rule does not bend. Truthful observation is one thing. Maintaining it is another. Keeping it, another still. The rule we apply to the system's memory applies to our own operations: the record never lies.

Where we are honestly

MARSI is an early company. We do not hold formal certifications yet, and we will not claim one before it is real. When we pursue and obtain SOC 2, ISO 27001, or equivalents, they will be listed here — with receipts.

Responsible disclosure

If you believe you have found a vulnerability in anything we operate, write to with the subject line SECURITY. We acknowledge reports within 72 hours, we will not pursue legal action against good-faith research, and we ask in return that you avoid data exfiltration and service disruption, and give us reasonable time to fix what you found before disclosing it publicly.

Contact: · Subject: SECURITY
← Back to marsi.ai

The constitution is live

MARSI runs under a ratified constitution today — not a policy document waiting to be enforced, but the operating condition of the system. Every rule, every change, and every exception traces to a ratified decision. The system audits itself, detects its own drift, and repairs it. Even the constitution can change only through its own transparent process.

What that means for anyone who relies on it

Nothing changes silently. Exceptions are recorded, not absorbed. Decisions carry their reasons, and reversals carry theirs. When the system tells you why something is the way it is, that answer is not a reconstruction after the fact — it is the record itself.

The human stays the source of meaning

Governance in MARSI orders how truth travels; it does not replace the people it serves. Lived experience is what becomes knowledge. Machines carry information — authority over meaning stays human.

What we will publish

As parts of the governance record become public-ready, we will publish summaries here with verifiable receipts, consistent with our Public Boundary. What we publish will state outcomes and their evidence — never mechanisms.

Status: constitution ratified · self-audit active
← Back to marsi.ai

The line, drawn deliberately

Everything MARSI publishes sits on one side of a line we drew before we published anything. On the public side: our philosophy, our stances, and our outcomes — what the system does, demonstrated as behavior. On the protected side: method — how it does it. Architecture, mechanisms, and internal structure are not published, referenced, or hinted at.

Why the boundary exists

Three reasons. The integrity of the work: our claims should stand on demonstrated behavior, not on persuasion about internals. The protection of foundational work while it matures. And discipline: a system whose whole thesis is accountability should be judged by how it behaves under scrutiny, not by how compelling its diagrams are.

The test we apply

Before anything becomes public we ask one question: does this reveal what the system does, or how it does it? Outcomes pass. Mechanisms don't. Where an outcome can carry a verifiable receipt, it does.

The boundary is not an excuse

If a claim we make publicly cannot be verified from what we have published, tell us: . We will either show the receipt or withdraw the claim. That is the deal that makes the boundary honest.

Test: what, not how · Receipts where verifiable
← Back to marsi.ai

MARSI will expose governed interfaces for connecting external agents and tools, including support for the Model Context Protocol. Documentation will be published here when the interface is stable enough to be worth your time — and not before. If you are building agent systems and want to be notified, write to .

Status: interface under development
← Back to marsi.ai

Acceptance

By using marsi.ai you accept these terms. If you do not accept them, do not use the site. We may update these terms; the date above changes when we do, and material changes will be noted plainly rather than slipped in quietly.

What this site is

marsi.ai is an informational site describing MARSI and its early-access program. Nothing here is a warranty of future functionality, an offer of service beyond what is expressly stated, or professional advice of any kind.

Use of content

The content of this site — text, design, and imagery — belongs to MARSI unless otherwise attributed. You may quote it with attribution. You may not use the contents of this site to train machine-learning systems without our written permission.

Early access

Participation in the early-access program is by invitation, may change or end as the work evolves, and is governed by the agreement provided at the time of access. We will tell you the truth about the system's status; in exchange, early access is provided as-is.

Liability

To the maximum extent permitted by law, MARSI is not liable for indirect, incidental, or consequential damages arising from use of this site. Some jurisdictions do not allow these limitations, in which case they apply to the fullest extent permitted.

Governing law

These terms are governed by the laws of the Commonwealth of Virginia, United States of America, excluding its conflict-of-law rules.

Questions:
← Back to marsi.ai

What we collect

If you write to us or request early access, we keep your email address and what you sent us. That is currently the whole list. This site sets no cookies and runs no third-party trackers or advertising scripts. If that ever changes, this page will say so before it does.

Whose data it is

Ours is the easier privacy policy to keep, because we mean the premise: your data is not ours. Your identity, your habits, your fingerprint belong to you, and a company built on provenance has no business holding what it cannot show a reason for holding. We don't want it. Trust is earned, never commanded — and this page is one of the places we earn it.

What your browser talks to

Serving this site may involve your browser fetching fonts, one public-domain image of Mars, and — only when you open Field Notes — GitHub's public API, so you can verify our receipts yourself instead of trusting our copy. Each request is the minimum the page needs to do its job; none of them carries anything about you from us; and we keep pulling these in-house as the site matures. This page will always state the list exactly.

How we use it

To reply to you, to evaluate early-access requests, and to keep an honest record of our correspondence. We do not sell personal data. We do not share it with third parties except where required by law or where a service provider processes it on our behalf under confidentiality obligations.

Retention and your rights

We keep correspondence as long as it is relevant to an ongoing relationship, then delete it. You can ask us at any time what we hold about you, ask us to correct it, or ask us to delete it — write to and we will act on it without ceremony. Depending on where you live, these rights may also be guaranteed to you by law (for example under GDPR or CCPA); nothing here limits them.

A note in our own voice

MARSI exists because records should be truthful and provenance should be total. A privacy policy that hedges would be beneath the point of the company. If you ever find this page and our practice out of alignment, tell us — that is a report we want.

Contact: · Subject: PRIVACY
← Back to marsi.ai

Why field notes

MARSI's thesis is that intelligence fails without a missing layer — provenance, shared context, honest revision. A thesis like that should not ask to be believed. It should point at the world and say: watch. Each entry here is a live, observable experiment — a claim about how intelligence breaks without the layer, stated plainly, carried by public receipts, and updated when reality moves. Entries revise; they never quietly vanish.

Experiment 001 — The wrong pool

Anthropic builds the best AI models in the world. We say that as admirers and daily users — this entry exists because we hold them in high regard, not in spite of it. What follows is not a story about carelessness. It is a story about a law that applies to everyone, including us.

The observation

Signing in to Claude is passwordless — an email code, or a Google or Apple identity. More secure in the abstract: nothing to phish, nothing to leak. But a password manager entry is really an identity record, and with no password to save, nothing ever gets written down. A person with three email identities now plays account roulette at every login screen. And the failure mode is sneaky: you don't get "wrong password." You get a perfectly working session in the wrong account — connectors, history, and settings that don't match what you remember. To you, the product looks broken. It isn't. You are simply not who you think you are — and nothing on the screen contrasts your belief against the system's record.

The mechanism

Why hasn't the world's best AI company seen it? Our inference — and we mark it as one, because that is the discipline this entry teaches: their engineers almost certainly sign in once and stay signed in. That daily fluency quietly filters this failure out of the very pool of problems they can observe. No amount of intelligence recovers a bug your own sampling excludes. The people best equipped to fix a problem are often the least likely to ever meet it. Their pool is shaped wrong — never their minds.

The consequence, on the record

The public record shows the scatter: independent reports describing the same phantom failure — a connector that says "connected" yet does nothing — each written from inside its own frame, none able to see the others, no shared truth to diff against. Read a few in Anthropic's own open issue tracker: #61682, #57589, #23775, #32479 — and #40238, filed because switching the linked account required revoking access from GitHub's side, then closed as a duplicate. The scatter, again. Nobody in those threads lacks intelligence. They lack a fixed point.

The states you see beside those links are not written into this page. When you opened this entry, your own browser asked GitHub's public API for the current state of each report. What you are reading is the record as it stands right now — including anything that has closed since we wrote this.

The lesson

Contrast is not a nice-to-have on top of shared truth — it is the payoff. You cannot detect drift without a fixed point. The system knew exactly which account was signed in; it kept that truth private, so a wrong belief ran unopposed. One contrast surface — you are signed in as X, by method Y — shown at the moment of confusion would dissolve the entire class of failure.

And the deeper lesson: the boundary is where governance must live. Every input arriving from outside a domain — a bug report, a login, a grant of access — carries its own coordinate system. Let it in raw, and each import fragments the shared truth into forty threads that cannot see each other. The inbox must be regulated from within the domain: not by gatekeeping content, but by stamping every incoming claim against the domain's current truth, so what lands is a legible delta instead of free-floating noise. Challenge stays fully open. It simply arrives in terms the domain can compute against.

Our position

If the layer is missing at the top, it is missing everywhere — and it would be missing in us. That is why MARSI treats it as structure, not talent. When this experiment resolves — when the contrast surface exists, or the pool gets corrected — this entry will record it, with the receipt. That is the deal that keeps a field note honest.

Entry 001 · Status: observable · Method: public record
← Back to marsi.ai

How to reach us

Everything routes through one address: . Early-access questions, partnership and press inquiries, technical questions from labs and engineers, and reports of any kind all land with people who work on the system itself — there is no tier of support that hasn't touched it.

What to include

Say who you are, what you are trying to do, and — if something went wrong — what you expected against what happened. Screenshots and timestamps help. If your report concerns security, put SECURITY in the subject line and see our disclosure terms.

What to expect

We are a small team and we answer honestly rather than instantly: expect an acknowledgment within two business days, and a real answer rather than a template. If we don't know yet, we will say that — and tell you when we will.

· Acknowledgment within 2 business days
← Back to marsi.ai