the0racle · In development · R&D
It does not predict: it proposes, measures and leaves a trail
the0racle is a research platform for auditable agentic trading: a proposer proposes, a deterministic risk engine decides, with an absolute veto, and every decision is recorded. Today it runs in simulation only. It is not an investment service and it is not open to the public.
Read the notice What has been built
- An internal R&D project
- Simulation only: it does not trade real money
- Not investment advice
Notice
Before you read on: what the0racle is not
the0racle is an internal research and development project of Singular Beacon. It is not an investment service, it manages nobody's money and it is not open to the public: it takes no clients, users or contributions.
Nothing on this page is financial advice, an investment recommendation, or an offer or invitation to invest.
Today it runs in simulation only, on a synthetic market. It does not trade real money: that mode does not exist, either in its configuration or in its code.
We publish no results or returns here. A simulation on a synthetic market is no evidence of how a strategy would behave in a real one.
The problem
If an AI proposes, who decides?
The question the0racle investigates is not how to be right. It is how to govern a system in which an AI agent proposes trades: who has the final say, and how you prove afterwards what happened.
Proposing is not deciding
A language model can propose, but it should not have the final say. There has to be a fixed rule, one with no opinions, between the proposal and the order.
No record, no review
If a decision leaves no trail, it cannot be reviewed. And a trail that can be edited afterwards is no use either.
A simulation can mislead
A test that ignores what trading costs, or that leaves the AI inside the loop, proves nothing. You have to know what does not count as evidence.
What it is
The path of a decision
the0racle is the machine that brings order to that path. The strategy it includes is only a reference to exercise the machine, not an investment thesis.
Someone proposes
A proposer, which can be a reference strategy or a language model, puts a trade forward. It has no access to execution.
The risk engine approves or rejects
A deterministic engine evaluates each proposal against its limits and has an absolute veto. It allows no structure with an undefined maximum loss.
Only what is approved is executed, in simulation
Every order stems from an approved risk decision. Today that execution is simulated.
Everything is recorded
Every relevant action goes into a linked audit chain that can be verified and cannot be rewritten.
And it can be stopped
There is a kill switch, and loss-based stop criteria that halt the machine. Starting again is a manual decision.
Progress so far
What has been built so far
The test for calling this first version done is that the machine works, vetoes correctly and is traceable. Not returns.
A prior feasibility analysis
Before any code was written: literature and state of the art, other people's real experience, technical options and the regulatory and tax framework, with a log of the decisions still open.
A first local version
It runs on a machine of our own, on a reproducible synthetic market.
The risk engine, tested against adverse cases
Position sizing and the risk engine are complete and tested with adverse cases.
Historical simulation with costs
It accounts for spreads, commissions, slippage and margin. And it does not allow a language model inside the loop: a test like that does not count as evidence.
Audit, stop and daily report
An immutable audit log, a kill switch, a daily report and a local web panel to look at it all.
Real data, for analysis only
Real data can be loaded for analysis, and every data point states its source. The engine still runs on the synthetic market.
The AI proposer, off by default
The language-model proposer exists behind the same interface and is switched off by default.
Who it is for
Today, a lab of our own
For ourselves
It is an internal R&D project: it helps us learn how to govern AI agents with deterministic rules and a verifiable record.
If the approach interests you
If you work in risk, audit or the governance of AI agents and would like to compare notes on the approach, let's talk. We are not looking for clients or funds.
Current status
In development, and in simulation only
- StatusIn development: first local version
- ModeSimulation only, on a synthetic market
- Real moneyNo: that mode does not exist in the software
- Open to the publicNo
- Business decisionsOpen: what has been decided so far is provisional
Shall we talk about how to govern an agent?
If the technical approach interests you, or you would like to collaborate on the research, write to us. This is not an investment service, we give no advice and we accept no funds.