Infrastructure for autonomous capital

Guardrails that let
AI agents trade
without hallucinating.

RiskOS is a four-layer stack that constrains every decision an autonomous agent makes — from signal generation to trade execution. Invariant inputs, bounded allocations, hallucination scoring, auditable outputs.

How it works

Five stages. Zero hallucinations.

Market Data Input
AxonSignal Layer
Processing

5 invariant models run in parallel

Fractal
GBM
Neural
Macro
Geodesic
Versioned risk artefacts
TensorAllocation Engine

Constrained portfolio allocation

Budget limits
Refusal modes
Concentration caps
Bounded allocations
MyelinVerification Layer

Hallucination scoring & audit

Source cross-ref
Trust calibration
Divergence check
Confidence-calibrated trust scores
SynapseDecision Layer

Aggregates into trading plan

Signal merge
Audit trail
Replay guarantee
Bounded, auditable trading plan
Validated Trading Plan Output
LLM Hallucination Guard
!

Without RiskOS

LLM fabricates justifications for trades. No source verification, no confidence calibration, no audit trail.

With RiskOS

Every claim is cross-referenced against invariant source data. Myelin scores divergence, Synapse only executes plans that pass all verification gates.

Verification confidence94.2%
Live Trace
axn_run_001 → fractal.v0.9 → regime:Range confidence:0.91
axn_run_001 → gbm.v0.6 → dist_tag:tight q50:+0.003
§2 — The Stack

Introducing RiskOS_

A tiered guardrail stack that constrains AI agents at every decision boundary. Four layers of infrastructure — from invariant signals to validated trading plans — that give autonomous agents structure, auditability, and safety at every step.

Axon

Signal Layer

Produces invariant risk artefacts — versioned, replayable model outputs that form the foundation of every downstream decision. Five models run in parallel: fractal regime detection, geometric Brownian motion, neural pressure scoring, macro regime classification, and geodesic momentum vectors.

Hover to explore →

Axon

Produces invariant risk artefacts — versioned, replayable model outputs that form the foundation of every downstream decision. Five models run in parallel: fractal regime detection, geometric Brownian motion, neural pressure scoring, macro regime classification, and geodesic momentum vectors.

View documentation →

Tensor

Allocation Engine

Constrains capital allocation into opposing buy and sell baskets with hard budget limits. Every position is bounded by risk budgets, refusal modes activate when constraints are violated, and the engine guarantees that no single agent can exceed its mandate.

Hover to explore →

Tensor

Constrains capital allocation into opposing buy and sell baskets with hard budget limits. Every position is bounded by risk budgets, refusal modes activate when constraints are violated, and the engine guarantees that no single agent can exceed its mandate.

View documentation →

Myelin

Verification Layer

Scores every LLM output for hallucination risk using information-theoretic measures. Acts as an MCP and API auditor — cross-referencing agent claims against source data, flagging fabricated justifications, and producing confidence-calibrated trust scores.

Hover to explore →

Myelin

Scores every LLM output for hallucination risk using information-theoretic measures. Acts as an MCP and API auditor — cross-referencing agent claims against source data, flagging fabricated justifications, and producing confidence-calibrated trust scores.

View documentation →

Synapse

Decision Layer

Produces the final bounded, auditable trading plan. Aggregates signals from Axon, allocations from Tensor, and trust scores from Myelin into a single foresight document — fully traceable, deterministically replayable, and ready for execution.

Hover to explore →

Synapse

Produces the final bounded, auditable trading plan. Aggregates signals from Axon, allocations from Tensor, and trust scores from Myelin into a single foresight document — fully traceable, deterministically replayable, and ready for execution.

View documentation →
The difference

Unconstrained agents vs. RiskOS

Unconstrained agent
Random asset selection → fabricated justification → arbitrary sizing → no trace
1
Search 100s of assets randomly
No structured universe, no invariant signals
2
Fabricate justification
LLM hallucinates a narrative for a single pick
3
Arbitrary sizing
No constraint engine, no risk budget enforcement
4
No trace, no audit
Decision is opaque, unreplayable, unverifiable
Unauditable. Unreplayable. Unbounded.
Why Sylaris

Built for agents that manage real capital

Sub-millisecond latency

Cached inference results served from edge. Heavy GPU compute happens on our side — your agents get fast, pre-computed signals.

Interpretable outputs

Every signal comes with human-readable explanations. No black boxes. Agents can reason about why a score is what it is.

Deterministic replay

Every decision tree is hashed and versioned. Replay any trade decision from any point in time. Full institutional audit compliance.

Regime-aware intelligence

Fractal analysis detects when markets shift between linear and non-linear regimes — before traditional statistical models break down.

4
GPU-powered models
Running in parallel
<1ms
API latency
Cached inference
100%
Decision auditability
Cryptographic hashes
2
Markets covered
US & UK, expanding
Get started

Ship agents that institutions trust.

We're working with a small number of teams building autonomous capital systems. If your agents need real risk infrastructure — invariant signals, hallucination detection, bounded allocations — we should talk.

# No sales deck. We'll send API docs and a sandbox key.