Draft — counsel must review
This is a working template prepared for review by qualified counsel in the United States, India, the United Kingdom and the European Union. It is not yet in force and creates no rights or obligations until a final version is published.
In plain English
- AI researches and proposes. Plain, deterministic code decides whether any order may be sent — the model can never place one itself.
- Our model is retrained often but only released after beating the current version out of sample, a shadow run and human sign-off.
- It can be wrong. It abstains when unsure, it shows its evidence, and you can always ask for a human review.
- Whenever you talk to Nexus Assist, you are talking to an AI unless it says a person has joined.
The summary helps you navigate; the sections below are the document.
1Where we use AI
| Feature | What the AI does | Who decides |
|---|---|---|
| Research and ideas | Reads filings, results and news; drafts theses with citations; combines with quant signals | You (Copilot) or your signed policy (Autopilot), then the risk gate |
| Goal feasibility | Statistical simulation of possible outcomes (not a language model) | You |
| Daily report | Summarises what happened in plain language from audited records | Figures come from the ledger, not the model |
| Nexus Assist | Answers questions from our published help content; hands off to people | Our support team for account actions |
2The models
- Research model: an open-weights language model (base model configurable, currently about 8B parameters under an Apache-2.0 licence) fine-tuned by AlgoNexus. Releases are named
nexus-YYYY.MM.DD. - Quantitative models: momentum, volatility, factor and regime models with calibrated probabilities.
- Portfolio optimiser and risk gate: deterministic code, independent of the models, covered by automated tests.
- A model card for each release — training data sources, evaluation results, known limitations — is published with the release notes.
3Training data
Public and licensed data: regulatory filings, company results, regulator circulars, market data and news under licence, plus the recorded outcomes of our own past predictions. Customer data is used only in de-identified or aggregated form, under our Privacy Notice, and you can object. We do not train on your chats with Nexus Assist unless you opt in.
4Evaluation and release gates
- Out-of-sample tests against the current model and simple baselines, net of realistic costs, with no worse drawdown.
- Finance-specific evaluation: factual accuracy on filings, citation accuracy, calibration, refusal of out-of-scope requests.
- Shadow run on paper alongside production.
- Sign-off by a named human reviewer.
- Canary release to a small share of paper accounts, with automatic rollback on regression.
5Human oversight and your controls
- Copilot requires your approval for every order. Autopilot is bounded by limits only you can change.
- Pause and kill switch on every device. The engine also stops itself on stale data or model disagreement.
- Every decision records the model and policy versions used, so it can be explained and audited.
- You may contest a decision and request human review at support@algonexus.trade.
6Known limitations
- Language models can misstate facts or citations; we check citations automatically but cannot catch every error.
- Models learn from the past; they can fail in new market regimes.
- Coverage is thinner for small companies and for some UK and EU securities.
- Non-English sources are currently under-represented.
7Regulatory approach
- EU AI Act: we assess each AI system against the Act’s risk categories. Our current assessment is that our systems are not high-risk under Annex III, which we keep under review with counsel. We meet the transparency obligations of Article 50: people are told when they interact with an AI system and AI-generated text is identified as such. We also follow the AI literacy duty for staff.
- United Kingdom: we follow the government’s AI principles (safety, transparency, fairness, accountability, contestability) and FCA expectations including Consumer Duty.
- United States: we avoid deceptive claims about AI capability (FTC Act) and follow applicable state AI laws.
- India: we follow SEBI’s requirements on the use of AI by regulated entities, taking responsibility for outputs and data security. [verify current circulars]
8Reporting problems
If you see an AI output that is wrong, harmful or biased, use “Report” on the output or email support@algonexus.trade. We review reports, log incidents and include a summary in our annual transparency report.