Policy-bound AI trading and intelligence
Cortex is the agent runtime and policy engine of ORBITRA ONE™. Agents observe markets and portfolios, reason with a model ensemble and propose trade intents — but no intent becomes an order until Aegis has simulated it and the policy engine has authorized it. Every pass through the loop leaves an action receipt. The trader’s side of this experience is described under intelligence.
How it moves
The six Cortex stages — Observe, Reason, Propose, Simulate, Authorize and Execute — form a closed loop. Eight policy gates stand before execution: capital allocation, daily and lifetime loss limits, leverage ceiling, market allow and deny lists, session and expiry, human confirmation, approved data sources and the kill control. An intent moves on to Execute only when every gate is open; a closed gate sends it back with the reason.
The problem
Most trading bots are trusted by possession. They hold an API key with broad permissions, and their limits sit inside their own code. When the code is wrong, the model drifts or the key leaks, nothing outside the bot notices until the account does.
AI models add failure modes of their own: confident reasoning built on stale data, inputs from unvetted sources, behavior that shifts when markets enter a regime the model never saw. Afterwards there is rarely a record of what the agent knew, what it concluded or why it acted.
Cortex separates thinking from authority. Agents may research broadly, but execution passes through limits the agent cannot edit, enforced by components the agent does not control and recorded in receipts the owner can inspect. Automation, including AI-assisted automation, can still fail or behave unexpectedly; the architecture bounds the consequences.
Operating sequence
Every agent action — from a hedge adjustment to a new position — travels the same six-stage loop. No stage can be skipped, and a failure at any stage returns the intent to the agent with its reason.
The agent reads market and portfolio state: signed Prism market data, its owner’s Aegis graph and any approved external sources. Each input is logged with its provenance.
A model ensemble evaluates the situation. Each model’s view and version is recorded; disagreement lowers confidence and can stop a proposal before it forms.
The agent emits a trade intent: instrument, side, size, price limits, expiry and rationale. An intent is a request, not an order.
Aegis applies the intent to the portfolio graph and returns margin used, liquidation distance, loss-limit headroom and concentration after the trade.
The policy engine checks all eight controls. If the action crosses a confirmation threshold, it waits for the owner’s signed approval.
The authorized intent becomes an order signed under the agent’s VaultID credential. ApexMatch evaluates it at the Aegis gate once more before matching, and the receipt closes the loop.
Architecture
Cortex is built so that the component that thinks is never the component that authorizes.
Isolated environments in which agents run. A research workspace with broad read access is kept apart from the narrow execution channel through which intents leave.
Several models evaluate each decision independently. Their versions, outputs and agreement are recorded, so every proposal can be traced to the models that produced it.
Evaluates each intent against the owner’s policy, written as versioned policy-as-code. Anything not explicitly permitted is denied.
Each agent holds its own VaultID credential, bound to its owner and scoped by policy. Agents never sign with the owner’s keys.
Connects every intent to the owner’s Aegis graph, so authorization is decided on post-trade risk rather than on the intent alone.
Writes a signed action receipt for every loop pass, rejected ones included: reason, simulated risk, provenance, policy verdict and execution evidence.
Security and failure control
The policy engine checks every control before an intent is signed, and the gate inside ApexMatch checks the credential’s limits again before matching. An agent whose own code fails still cannot exceed its authority. The automated and AI-assisted trading disclosure explains the risks that remain.
Across the three systems
Orbitra Prime
Trading intelligence
In Orbitra Prime, Cortex proposals appear in the guidance panel with their reason, simulated risk and data provenance, while confirmations and the kill command are reachable from desktop and mobile alike. Manual and agent orders share one account, one risk view and one evidence trail.
Orbitra L1
Settlement and compute
Agent credentials and policies are state on Orbitra L1, so enforcement never depends on the machine an agent runs on. Agent orders pass the same Aegis gate inside ApexMatch as manual orders, and receipts reference finalized transitions.
Orbitra Realm
Applications and commerce
In Orbitra Realm, agents are first-class participants with their own identity: they can consume data from the data exchange, pay for compute and be offered through the bot marketplace — always under an owner’s policy. See the AI economy.
Value
Specifications
Terminology