AI
NOVAQ uses AI in three roles: as an analyst of the market, as a builder and auditor of the system, and as a subject of research. Each system below carries its actual status.
What AI does at NOVAQ
| System | What it does | Status |
|---|---|---|
| Intraday analysis agent | Writes structured market briefs from verified fields, about 93 scheduled slots a trading day | Running, no order rights |
| Execution and parity agents | Two agents check executions and parity with the system. They are authorised to intervene in emergencies: collisions, unprotected positions, unexpected situations and capital protection | Active around the clock |
| Audit swarm | Eight specialised roles audit research and running code | Active, read-only by default |
| Market context expert | Classifies market state from data and distilled vault knowledge | Shadow |
| AI trader | Autonomous trading experiment under the agent gate | Simulation |
| Neural models | Estimate price behaviour | Simulation, not validated |
| Probabilistic sizing agent | Scales exposure | Simulation |
| Neural trade filter | Gives a second opinion on signals | Simulation |
| Council and supervisors | Aggregate expert views | Shadow |
| Further AI experts | Author-specific analysis | Disabled |
| Local language model | Runs on own hardware | Disabled after a month in shadow |
| Agent Arena | Competition of AI participants | Planned, design documents only |
Statuses were last checked in October 2026.
An analyst that works from verified data
Capability. During each session, the agent assembles a multi-layer picture of the market and writes a brief.
What it enables. A reader gets a consistent narrative tied to specific fields, so each statement can be traced to its source.
Result. All 253 fields were independently recomputed from raw data. The first pass confirmed 202 and flagged the rest, three rounds of fixes followed, and a live audit then covered 60 bundles.
Background. Lessons from this work moved meaning into the data layer: each field now states its own direction and role, so the model does not have to infer them.
A knowledge-based expert
Capability. The market context expert reads live data, session statistics, the economic calendar and a distilled description of market states.
What it enables. Twelve universal and seven session states from the vault reach a running expert. Each version is hashed into its input record, so any change in what the expert knows can be audited.
Background. It runs in shadow and can decline to give a view. The first run did exactly that because the data was not fresh. Experts are isolated by author through provenance filters, which keep one author's concepts out of another's analysis.
An AI audit swarm
Capability. Eight specialised roles cover orchestration, red team, strategy, account rules, simulation, execution, development and quality control.
What it enables. Independent checks on research and live code, with each finding carrying an author and a second signature, and fixes needing separate sign-off.
Result. Since 5 September, 758 runs, 311 verification files and about 100 numbered findings. The swarm traced 1,707 relevant commits and cross-linked 598 defect records from seven sources. It found failures that raise no error, including a family of time-zone problems and skipped signals that left no trace.
Eight roles around one recorded ledger
Multi-agent research and adversarial review
One model builds and another tries to prove it wrong. In July 2026 a formal loop ran 20 rounds in one day under strict rules: production data read-only, and no agent's statement accepted as its own proof. The same principle now runs through the audit swarm and large research studies.
The AI trader
The AI trader is a live experiment on a simulation account, with a full record of each decision. It went through a formal trial against criteria for net result, sample, drawdown, profit concentration and rule compliance.
The verdict on 12 October 2026 was edge not demonstrated, with a recommendation not to deploy real capital. The trial's conditions changed along the way, so it reads as a verdict and not as a clean test. Its value lies in the measurement: for example, a counterfactual on 203 trades showed that active trade management had no effect distinguishable from zero, which pointed the work at entry decisions. The experiment continues with simpler, measurable rules.
Neural research
Neural models are a research track at NOVAQ. They run only on simulation accounts and are not counted as validated, and this site publishes no separate neural performance figures.
Evaluating language models under production conditions
NOVAQ tested 10 models with byte-identical production prompts and production validation. Five met the output contract in both runs. A separate local model then ran for a month in shadow, producing 1,567 records with 99.9% format compliance, and it was disabled because latency stayed above the 90-second limit. Model choices rest on measured behaviour. Each cell of the 10-model test had one run, so that comparison is indicative.
AI in development
NOVAQ was built by one operator working with AI coding agents. The operator sets direction, makes decisions and judges results. Agents write code, run research, produce documentation and supervise processes. In under six months this produced a platform, a research laboratory, a knowledge base and an audit system, and the audit layers keep that speed under control.