Division

A paper-first trading experiment.

Division explores whether evidence-linked AI decisions can produce repeatable net profit after trading and AI costs, within explicit loss limits. Profit is the objective—not a promised result.

Evidence first /
authority stays bounded

This map describes the intended system, not a claim that the connected paper beta is running. The local components have been checked with controlled sources and simulated execution.

Facts → proposals

Time-correct market facts, account state and costs inform Pulse, Echo and Trace. Each proposes an evidence-linked action. Entitled real-data integration still needs qualification.

Local engine

Challenge → selection

Atlas challenges evidence and risk. Switch selects only an original eligible proposal—or rejects every proposal. It cannot invent a new trade.

Five AI roles

Fixed checks → durable intent

Independent software checks sizing, cash, exposure and loss limits. A durable intent records the action; uncertain outcomes must be reconciled, not blindly retried.

Software enforcement

Paper adapter → journal → facts

The intended Moomoo paper path records confirmed orders, fills and account changes, then feeds that journal back into the next decision. The connected broker lifecycle and actual paper account remain pending qualification.

Connection pending

Replay → comparison

Causal replay and independent comparison books measure net returns, drawdown, costs, coverage and latency. Synthetic results cannot establish a trading advantage.

Evaluation, not execution

Public explanation /
private application

This page is not a trading application. The separate Division app is designed for sole-owner access, not public signup.

The first hosted release is sample-data only: Common owner login → Floor, Crew, Taxes and audit views. Common authenticates the owner; Division’s separate model connection supplies inference. Those are different connections.

The hosted-to-local bridge is pending, not implemented. The intended future path is a qualified local journal → private website connection. The first hosted website receives no local broker credentials, actual account records or trading control. The separate app address is division.ryoi.ai; this page does not change its access gate or establish that its hosted login is live.

Replace the mechanism /
not the safeguards

Models and personalities are means to the financial objective. A faster selector is not automatically a better investment decision.

Responses baseline

GPT-6 Luna through the OpenAI Responses API completed one structured five-role round on fabricated facts. That verifies transport and the original role, evidence and selection contract—not market integration, reliable trading or profitability.

Synthetic proof only

Decisions inside Switch

The OpenAI Decisions trial is pending. Its proposed place is inside Switch’s finite selection—not a sixth manager or extra veto. Any replacement must preserve supported reasons, evidence references and treatment of Atlas’s objections; a typed choice alone is insufficient.

Trial pending

Comparison without authority

TypeSafe Jev is a comparison candidate only. Shadow evaluation has no order capability. Probabilities must never be presented as invented model reasons or allowed to change financial authority or fixed sizing limits. There is no measured performance winner.

Not a production dependency

What exists /
what remains

Architecture snapshot: 17 October 2026. Evidence levels are separate: implementation, connection proof, paper performance and real-money authority are not interchangeable.

Controlled implementation

Crew, risk checks, ledger and replay have controlled-source and simulated-execution verification. The website has working local sample-data views; its private Common-login implementation is prepared. This does not prove a connected hosted trading system.

Implemented in isolation

Connected paper path

Entitled real data, a coherent actual paper account, broker lifecycle proof and the hosted-to-local bridge remain to be qualified. The optional selector experiment does not postpone the genuine paper beta.

Pending

No real-money activation

Real-money operation is not armed. It requires paper evaluation, separate exact financial authorization and bounded live proof. No net-return or profitability result is claimed here.

Not authorized by this page

Let’s make the idea real

Contact

Have a product idea, an opportunity to collaborate, or something worth bringing back to life?

DivisionDivision

Evidence first /authority stays bounded