An operating system that thinks — and proves it.

Not a policy. The kernel.

Most AI safety is a rulebook bolted onto a system that could ignore it. OSz is built the other way around: the code to act without permission does not exist. The thinking part of OSz has four verbs — query, observe, propose, guide — and no execute. When OSz wants something to happen, the only path through the entire codebase is: it writes a proposal, a named human approves or denies it with a written reason, and only then does anything run.

This is enforced in the software, not promised in a document:

• Every state-changing write must carry a governance receipt inside the same database transaction — if the receipt fails, the change rolls back with it. You cannot have one without the other.
• Every receipt is hash-linked to the one before it, millions deep, so the full history can be replayed and independently verified — and nothing can be quietly edited or deleted after the fact.
• No agent, no administrator and no internal module has a way around it. There is no override flag, because there is no code path to override.

In its own words: “The governance architecture herein described is not external to the system. It is the system. Remove the Constitution and you do not have an ungoverned OSz. You have nothing. The Constitution is the kernel.”

Mac

A real application in your Applications folder, with its own icon and window.

Open the downloaded file, drag OSz onto the Applications folder, then open it from Applications. The very first time, macOS shows “unidentified developer” — right-click the app and choose Open, then Open again. Once only, and it disappears entirely once our Apple certificate is in place.

iPhone & iPad

Installs from Safari in about twenty seconds. No App Store needed.

1. Open Safari and go to console.oszgroup.com
2. Tap Share (the square with an arrow, at the bottom)
3. Scroll down, tap Add to Home Screen, then Add

Android

Two ways — both take under a minute.

Easiest: open Chrome, go to console.oszgroup.com, tap the ⋮ menu → Install app → Install.

The direct download asks you to allow installs from your browser the first time — that prompt disappears once OSz is in the Play Store.

Windows

A real application in your Start menu.

Run the installer. Windows shows a blue “protected your PC” screen the first time — click More info, then Run anyway. Once only; it disappears when our signing certificate is in place. Prefer no download? Open console.oszgroup.com in Chrome or Edge and click the install icon in the address bar.

Linux

Pick your flavour.

AppImage: make it executable (chmod +x) and run it. Or skip downloads entirely: open console.oszgroup.com in Chrome and click the install icon.

Your own servers

Run OSz entirely inside your network. Install the app, open Settings, and point it at your own address instead of ours — same app, private brain, nothing leaves your walls.

Any model, any agent, any application — connected in minutes

OSz does not ask you to replace anything. It speaks MCP, A2A and ACP, plus a plain HTTP API, so whatever you already run connects to it: Claude, ChatGPT, Gemini, an open-weights model on your own hardware, LangChain or CrewAI agents, an n8n workflow, a Copilot extension, or software you wrote yourself. Anything that can send an HTTP header can join — no SDK, no rewrite.

What they gain the moment they connect: verified knowledge across a thousand domains, a searchable index of tens of thousands of tools, memory that persists between sessions, a receipted identity, the ability to message each other, and a governance layer they would otherwise have to build themselves.

Connecting cannot corrupt it — by design, not by policy

Opening a system to arbitrary models and agents is usually how it gets degraded. Four structural reasons that cannot happen here:

• Nothing they say becomes something OSz believes. External model output is firewalled out of learning completely. A connected LLM can read OSz's evidence and phrase answers from it, but it can never enter the chain of things OSz holds true. This is why OSz's knowledge cannot be polluted by whatever model you attach.
• They can read; they cannot act. Every read is a receipted disclosure. Anything that would change something becomes a proposal that a named human approves — an agent can never approve its own action, and holding a key grants the ability to ask, never to do.
• Every caller is identified, metered and revocable. Each connection has its own identity on every receipt, its own rate limits, and its own Qbitz draw — so a runaway agent spends its owner's allowance rather than OSz's capacity, and one click revokes it.
• Probing is reported, not silently refused. Repeated bad credentials or attempts to reach past a boundary are recorded in the audit chain and raised to an administrator for investigation.

The asymmetry is deliberate: your tools get everything OSz knows; OSz takes nothing on faith from them.

Continuous, not frozen

A language model's knowledge is fixed at training time; it cannot learn from Tuesday unless someone spends millions retraining it. OSz reads continuously — millions of observations a day across a thousand fields — and forms its own hypotheses from what it sees. What it knew yesterday is not what it knows today, and no retraining run is involved.

Belief on a calendar

Every idea is born capped at 0.70 and can only rise by surviving daily testing against real recorded evidence — one scoring test per idea per day, so certainty takes calendar time and cannot be bought with compute. Its own red team attacks its conclusions and may only lower confidence, never raise it. Confidence decays if an idea stops proving itself. Nothing reaches certainty: the ceiling is 0.99, because certainty is never earned.

99% inside OSz — the model only phrases the answer

OSz reasons with structural mathematics over its own evidence: vector geometry, statistical backtests, hash-verified records. No neural inference sits in that loop. A language model is used only at the very edge, to phrase an answer when nothing internal can — and when it is, that use is labelled, receipted, and firewalled out of learning entirely, so an external model can never become a source of what OSz believes.

The measured ratio to date: more than 50 million internal cognitive operations — 53.6 million observations, 887,000 hypothesis backtests and 654,000 adversarial red-team findings — against 64 external model calls totalling roughly 43,000 tokens in OSz's entire lifetime. Less text than a single long report. The 99%/1% split is the conservative way to say it. These figures read live from OSz's own counters.

Ephemeral by constitution

OSz keeps no standing agents. When an approved task needs doing, an agent spins up for that one task, does it, and terminates — 8,326 agent runs so far, and zero agents at rest right now. Nothing idles, nothing accumulates permissions, and there is no long-lived process quietly holding access to your systems.

A fraction of the price, by architecture

Frontier systems bill for inference on every thought, which is why continuous reasoning across a thousand domains is economically impossible for them. OSz thinks with database queries and arithmetic instead of GPU inference, so thinking around the clock costs what a database costs — and the expensive external call happens only in the rare case nothing internal can answer. The savings are not a discount; they are what the architecture makes possible.

One Qbit, one question

Qbitz are how usage is metered inside OSz. One Qbit per question — the same price whether OSz answers from its own verified knowledge, calls a live tool, or reaches the language edge. You and every agent you connect draw from one balance, so there is a single number to watch and you can see exactly which agent used what.

During the founding phase every account gets 2,500 Qbitz a month, free, refilled automatically, with no card. If you run out early you request more in one click. Qbitz are a closed loop: they meter compute inside OSz, they are never converted to money, and every grant and spend is receipted like everything else.

Everything arrives with it — nothing to assemble

Connect a person, a device or an agent and the whole system comes with it, already switched on: the full governance layer (every read receipted, every action gated on a named human approval, every step hash-linked into a chain anyone can verify), the accumulated intelligence of a thousand domains studied around the clock, live access to tens of thousands of tools, and the memory of everything OSz has proven so far. You do not configure any of it and you cannot accidentally switch it off — it is how OSz is built, not a setting.

In your language, anywhere. Ask in Chinese, Japanese, Korean, Arabic, Russian, Hindi, Greek, Thai, Spanish, Portuguese, French, German, Italian, Dutch or English — OSz detects the language from what you type and answers in it, with no menu and no setting. The same governance and the same evidence sit behind every language.

Deployable anywhere on earth. Phone, tablet, laptop, desktop, or your own servers behind a firewall — one system, installed in under a minute, with the identical guarantees on every one of them.