04Artificial Intelligence

Intelligence with a mandate.

We are designing autonomous agents the way institutions design roles: each with a defined mandate, explicit limits and a record of every decision it makes.

Architecture

Agents, a digital twin and an independent supervisor.

Autonomous agents and digital twinConceptual architecture. People set the mandate. Research agents feed a strategy that decides within the mandate; execution runs in simulation only and the live market connection is not active. A digital twin mirrors the strategy state. Supervision agents run on independent infrastructure, observe the strategy, the twin and execution, and can halt execution. Record agents keep an ordered, tamper-evident audit trail of every event.MandateObjectives, limits, permissions · set by peopleResearch agentsData, hypotheses, evidenceStrategyDecides within the mandateExecutionSimulation onlyDigital twinIndependent replica of stateMarket connectionNot activeSupervision agentsIndependent infrastructureOBSERVES STRATEGY,TWIN AND EXECUTIONCAN HALT EXECUTIONRecord agentsOrdered, tamper-evident audit trail of every eventLIMITSMIRRORS STATENOT CONNECTEDHALTOperationalIn developmentPlanned
Autonomous agents and digital twinConceptual architecture. People set the mandate. Research agents feed a strategy that decides within the mandate; execution runs in simulation only and the live market connection is not active. A digital twin mirrors the strategy state. Supervision agents run on independent infrastructure, observe the strategy, the twin and execution, and can halt execution. Record agents keep an ordered, tamper-evident audit trail of every event.MandateSet by peopleResearch agentsData, hypotheses, evidenceStrategyDecides within the mandateDigital twinMirrors strategy stateExecutionSimulation onlyMarket connectionNot activeINDEPENDENT SUPERVISION · CAN HALTRecord agentsTamper-evident audit trailOperationalIn developmentPlanned
People set the mandate. Execution exists only in simulation; the live market connection is not active. Conceptual architecture · status as of October 2026

Autonomous agents

Specialized roles, not a single black box.

Research

Research agents

Designed to monitor data, generate candidate hypotheses and prepare evidence for human review.

Risk

Supervision agents

Designed to watch exposures and limits continuously, and to halt activity without depending on the systems they supervise.

Audit

Record agents

Designed to keep a complete, ordered record of every proposal, decision and change, so any outcome can be reconstructed.

Digital twin

Every position, mirrored.

Our architecture is designed to pair each strategy with a digital twin: an independent replica of its expected state, used to detect divergence between what a system should be doing and what it is doing.

When the two diverge beyond a defined tolerance, supervision intervenes. The twin is how we intend to make autonomy observable.

Principles

How intelligence is allowed to operate.

Models as instruments, not oracles

Predictive models should produce calibrated probabilities with stated uncertainty. They inform decisions; they do not replace the decision process.

Humans set the mandate

People define objectives, limits and permissions. Agents operate within them and escalate whatever falls outside.

Every action is explainable

An agent that cannot show the evidence behind an action does not take it.

Supervision is independent

The system that supervises is never the system that acts.