How we use AI.
Where we don't.
This page explains - in plain language - which parts of TradeMire use AI, which don't, what data reaches which model, and the rails we've built to keep it safe. Written for traders, compliance teams, and regulators.
Algorithmic trading is not risk-free.
Crypto-assets are volatile and may lose significant value. Algorithmic trading can amplify both gains and losses. Limit orders may not fill; market orders may execute with slippage. Trading via TradeMire is non-custodial - your funds stay on your exchange or in your wallet. You alone are responsible for the orders you submit.
1. Scope
This policy covers every production system at TradeMire where a machine-learning model - large language model, classifier, forecaster, or embedding - contributes to a user-visible decision. It does not cover deterministic rule engines or static statistical models, which we document separately in the engineering handbook.
AI writes drafts and surfaces suggestions. It does not place unsupervised trades with live capital. Every live order requires a human-authorised policy signed by the account owner.
2. Our four principles
Four non-negotiable design principles - written down, referenced in every RFC, blocking at code review.
Auditable by default
Every AI output is logged with model version, prompt, input hash, and response. Retained 2 years (Pro) or 7 years (Institutional).
Human-authored policy
AI can draft a strategy, but a human signs the policy before it can touch live capital. No auto-promotion from paper to live.
Refusal over speculation
Models are prompted to refuse rather than invent. Fabricated numbers, made-up tickers, or hallucinated order types are treated as a Sev-1 bug.
User data is never training data
Your trade history, strategies, and prompts are not used to train any model. Ever. Contractually enforced with providers.
3. The models we use
Only the models listed below may run in production. Any additional model requires a signed model card, risk assessment, and security review.
| Model | Provider | Where | What it does |
|---|---|---|---|
| Claude Sonnet 4.6 | Anthropic | EU region | Agent chat, strategy drafting, research summarization, code assist. |
| Claude Haiku 4.5 | Anthropic | EU region | In-product explanations, tooltip expansions, fast classification. |
| text-embedding-3-large | OpenAI | EU region | Strategy search, research corpus indexing. No raw trade data. |
| TradeMire-Signals-v4 | TradeMire (internal) | EU - self-hosted | Order-book feature extraction. Trained on public market data only. |
| TradeMire-Risk-v2 | TradeMire (internal) | EU - self-hosted | Real-time position-risk scoring. Deterministic gradient-boosted trees. |
4. Data flows
The short version of what leaves your browser, what stays with TradeMire, and what reaches a third-party model.
- Your prompt to the Agent - sent to Anthropic (EU region) with zero-retention flag enabled. Not used for training under our commercial agreement.
- Your vault balances, order history, PnL - never forwarded to any external model. The Agent receives only a redacted snapshot you explicitly approve.
- Your API keys - encrypted at rest with a customer-specific KMS key. Never transmitted to any model, internal or external.
- Public market data (orderbook, OHLC) - used to train our internal signal models. Your activity is not part of this training set.
5. Human oversight
Human oversight is fundamental to our AI governance. The Agent can suggest, draft, explain, and simulate. It cannot act autonomously on live capital.
Human-in-the-Loop Controls
All automated trading strategies include configurable intervention thresholds. Users can set stop-loss limits, position size caps, drawdown triggers, and time-based restrictions to maintain meaningful control over algorithmic decisions.
Override & Kill Switch
Every active strategy includes an immediate termination capability. Users can pause, modify, or terminate any AI-driven process instantly through the dashboard or mobile interface.
Escalation Protocols
Anomalous market conditions, unusual model behavior, or threshold breaches automatically escalate to human operators. Critical events page on-call engineers within 60 seconds.
Interpretability & Explainability
Trading signals and analytics outputs include explanation layers that describe the key factors driving each recommendation. Black-box decisions are not deployed in user-facing flows.
What the Agent cannot do
- Place an order against live capital without an explicit policy signed by the account owner.
- Modify risk limits, kill-switches, or permissioning.
- Connect new venues, rotate API keys, or change account roles.
- Initiate or authorize withdrawals at any layer - the platform itself does not hold that capability.
Institutional customers can further restrict the Agent to read-only mode across all surfaces, organization-wide.
6. Evaluation & testing
Every model in production has a published eval suite, run weekly and on every prompt or weights change. Results are retained and available to Institutional customers on request.
- Hallucination rate on a 400-item financial QA set (target: <0.5%).
- Refusal correctness on adversarial prompt panel (target: >98%).
- Latency P50 & P99 tracked against SLAs.
- Red-team rotation - two weeks per quarter, external firm.
7. Your rights & controls
Every customer - not just Institutional - has full control over how AI features apply to their account. The EU AI Act grants additional rights to individuals affected by AI decisions; we honour them globally.
Controls available to every account
- Disable the Agent globally from account settings.
- Export the full log of AI interactions in CSV or JSON.
- Request deletion of stored AI context for any closed strategy.
- Opt out of all in-product recommendations.
Rights granted under the EU AI Act and GDPR
To exercise any of these rights, email ai-compliance@trademire.ai.
8. EU AI Act compliance
The EU AI Act is the world's first comprehensive AI law. We track its phased rollout and publish our compliance status as we go. We don't claim certification we don't hold.
Alignment statement
TradeMire's AI surfaces are aligned with the applicable obligations of Regulation (EU) 2024/1689 ("EU AI Act"). The risk pyramid below shows where each surface sits within the Act's classification framework. We track the rollout's effective dates and update this page as obligations come into force - we don't claim certifications we don't hold.
Regulatory Overview
The EU AI Act establishes a comprehensive legal framework for artificial intelligence systems in the European Union. As an algorithmic trading platform serving European customers, TradeMire is committed to full compliance with this regulation.
EU AI Act
Regulation (EU) 2024/1689 - the world's first comprehensive AI law, establishing harmonized rules for AI systems across the European Union.
Effective Timeline
Phased enforcement: prohibited practices (Feb 2025), GPAI rules (Aug 2025), high-risk obligations (Aug 2026), full enforcement (Aug 2027).
Scope of Application
Applies to providers and deployers of AI systems placed on the EU market or whose output is used within the EU, regardless of where the provider is established.
Risk Classification
The EU AI Act classifies AI systems into risk categories. TradeMire's algorithmic systems are assessed under this framework with appropriate safeguards applied to each tier.
Prohibited AI practices including social scoring, real-time biometric surveillance, and manipulative systems. TradeMire does not deploy any such systems.
AI systems in critical sectors (credit scoring, insurance, essential services). Our trading algorithms are assessed against high-risk criteria and meet applicable obligations where relevant.
AI systems with transparency obligations - users must be informed they are interacting with AI. Our analytics and signal-generation features include clear AI disclosures.
AI systems posing minimal or no risk. General-purpose analytics dashboards and visualization tools fall into this category.
Technical & Organizational Measures
We implement comprehensive technical and organizational safeguards to ensure our AI systems meet the requirements of the EU AI Act and align with global best practices.
Risk Management System
A continuous, iterative risk management process identifies, analyzes, evaluates, and mitigates risks associated with each AI system across its lifecycle.
Data Governance
Training, validation, and testing datasets undergo systematic quality assessment. We implement bias detection pipelines, data lineage tracking, and statistical representativeness checks.
Technical Documentation
Comprehensive technical documentation covers system architecture, algorithm design decisions, training methodologies, and performance characteristics for every AI component.
Record-Keeping & Logging
Automatic logging of AI system events enables traceability throughout the lifecycle. Decision logs, model versions, input snapshots, and output records are retained per policy.
Transparency & Information
Users are clearly informed when interacting with AI-generated outputs. Instructions for use, system capabilities, known limitations, and confidence intervals are surfaced in-product.
Accuracy, Robustness & Cybersecurity
AI systems are designed to achieve appropriate levels of accuracy, robustness against adversarial attacks, and resilience to data drift, with continuous monitoring.
AI Governance
Our governance posture matches the size and stage of the company. AI compliance is a designated responsibility within engineering leadership and is reviewed quarterly. We document, revise, and escalate as obligations come into force.
Designated responsibility
AI compliance is owned within engineering leadership. Regulatory developments are tracked, internal assessments are scheduled quarterly, and material updates are escalated to leadership.
Continuous monitoring
Production AI surfaces are monitored for performance, drift, and emerging risks. We re-assess and update model cards as obligations evolve or surfaces change.
Regulatory Timeline & Updates
The EU AI Act is being enforced in phases. We track our compliance progress against each milestone.
Prohibited Practices
Ban on unacceptable-risk AI systems takes effect.
AdherentGPAI & Governance
General-purpose AI model rules and governance obligations apply.
AdherentHigh-Risk Obligations
Full requirements for high-risk AI systems in Annex III become enforceable.
On TrackFull Enforcement
All remaining provisions, including Annex I high-risk systems, fully enforceable.
Preparing9. Global AI governance
While the EU AI Act is our primary compliance framework, we monitor and align with emerging AI regulations worldwide.
Our AI development follows the OECD's five principles: inclusive growth, human-centered values, transparency, robustness, and accountability.
We align our AI management system with the ISO/IEC 42001 standard for establishing, implementing, and improving AI management practices.
Risk assessments follow the NIST AI Risk Management Framework methodology: Govern, Map, Measure, and Manage.
Our crypto-specific AI tools are designed with Markets in Crypto-Assets Regulation (MiCA) requirements in mind.
10. Incident & disclosure
Material AI-related incidents - erroneous output that affected user decisions, model outages, provider changes - are logged on the Announcements page within 5 business days, with a full post-mortem.
Report suspected AI-related issues to security@trademire.ai with PGP-encrypted disclosure. A third-party bug bounty channel is in preparation. Acknowledgement within 24 hours.
11. Contact
AI Compliance Contact
For questions about our AI governance practices, compliance status, or to exercise your rights under the EU AI Act:
TradeMire - AI Compliance, Istanbul, Republic of Türkiye
National competent authority: KVKK (Kişisel Verileri Koruma Kurumu) and any AI supervisory authority designated under Turkish law. EU users may contact their national AI authority designated under the EU AI Act.