The Montrix AI Research Program treats AI trading as a governed decision system: models estimate, rules constrain, execution accounts for cost, supervisors control exceptions, and every decision remains reconstructable.
Research and implementation framework
The Montrix Research Program defines the architecture, validation standards, and implementation framework guiding the platform's ongoing development. Availability of individual capabilities depends on the activated product mode and deployment.
The papers are designed to be read together. Each addresses a different requirement of credible AI execution: how the system is structured, how its claims should be tested, and how model output becomes an admissible action.
Paper 01
18 pages · May 2026Architecture
Defines the Governed AI Execution Stack: a layered architecture connecting probabilistic market-state estimates, exposure proposals, deterministic risk constraints, cost-aware execution, supervision, and audit.
Read on ZenodoPaper 02
9 pages · June 2026Evidence protocol
Defines how governed AI agents should be evaluated across calibration, walk-forward testing, overfitting controls, benchmarks, ablations, stress scenarios, venue execution, and supervisor interventions.
Read on ZenodoPaper 03
6 pages · June 2026Control logic
Translates the architecture into implementation-ready logic for signal eligibility, exposure proposals, risk-envelope projection, venue-aware execution, supervisor escalation, and minimum audit records.
Read on ZenodoThe research does not claim that a model can forecast markets perfectly. It argues that AI trading becomes scientifically credible when model proposals are constrained, validated, cost-aware, supervised, and auditable.
/ 01
Forecasting and policy layers estimate state and propose exposure. Their output is not automatically an executable trade.
/ 02
Deterministic gates can reject or reduce proposals when cost, liquidity, venue health, drawdown, or event pressure make them unsuitable.
/ 03
A profitable signal can become unprofitable after fees, slippage, latency, funding, and adverse selection.
/ 04
Trades, rejected trades, no-trade decisions, and supervisor actions require sufficient records for later review.
The framework separates inference, risk, execution, and supervision so each layer can be tested independently and the complete decision can be reconstructed.
Check freshness, timestamp order, structure, latency, venue health, and account state before considering a signal.
Feature and signal layers estimate return, direction probability, volatility, drawdown risk, execution cost, liquidity stress, and event pressure.
A policy layer proposes target exposure but remains separated from the authority to execute.
The proposed target is constrained by leverage, concentration, drawdown, loss gates, liquidity, funding stress, venue limits, and supervisor state.
The execution layer selects venue and order type using expected slippage, latency, fill probability, rejection risk, and adverse selection.
Supervisor rules can block new risk, reduce exposure, request approval, switch venue, disable a module, or force reconciliation.
Every action and non-action records the relevant market, signal, policy, risk, execution, outcome, and supervision state.
The validation paper separates development evidence, launch-candidate evidence, and external-audit evidence. It asks whether attractive results survive realistic costs, model-selection controls, stress events, and independent reconstruction.
The framework does not treat supervision as an informal override. When conditions deteriorate, explicit states can restrict exposure, pause execution, require review, or force reconciliation. Accountability remains part of the trading logic.
Model sophistication matters, but it is insufficient by itself. Credibility depends on how forecasting, risk, execution, validation, supervision, and audit work together.
Anonymized validation evidence can support architecture, governance, calibration, stress-response, and risk-control claims. Public investment-performance claims require a higher standard: source order identifiers or verifiable hashes, account and custody records, capital assumptions, fee and funding records, venue-status logs, benchmark definitions, and independently reproducible calculations.
The evidence boundary is not a weakness. It separates internally validated research from externally audited claims and makes the research program more credible.
The objective is not to build a model that trades every signal. It is to develop an execution system that recognizes when a signal is admissible, when risk must be reduced, when execution quality is insufficient, when supervision is required, and how every decision can be audited.