CactusQuant: Making a Self-Improving AI Trading System Safe
CactusQuant · An Osonwanne Group LLC venture
A self-improving AI trading system, engineered to stay safe
CactusQuant is building an AI that manages a systematic trading strategy and improves itself over time. The patent-pending innovation is not a market prediction. It is the safety machinery that lets the system change itself with guardrails it cannot quietly route around.
Most conversations about AI in markets fixate on prediction: can a model see the next move before everyone else? At CactusQuant, an Osonwanne Group LLC venture, we found a more interesting and more durable problem to solve. CactusQuant is an artificial-intelligence system that manages a systematic trading strategy and improves itself over time, retraining its own decision policies, tuning its own settings, and learning from its own results as markets evolve.
The genuinely hard part is not getting an AI to change itself. The hard part is letting it change itself with capital at risk while engineering it so it can never endanger the account. That is the work we protected. CactusQuant is now patent-pending for the safety architecture behind a self-improving AI trading system.
The hard problem is not self-improvement. It is safety.
A system that can rewrite its own behavior is powerful and dangerous in equal measure. One mislabeled result, one faulty model, or one runaway tuning step can compound into real damage. So the question we set out to answer was not “how do we make the AI smarter?” It was “how do we let the AI get smarter without ever letting a single bad change put the account at risk?”
Our answer is a layered set of guardrails that govern every form of self-change. The filing covers those systems and the controls around them, not any proprietary formula and not a promise of returns. Below is the architecture at a benefit-level altitude: what the systems do, deliberately without the how.
The safety architecture, in plain terms
Policies that earn the keys
As markets shift, the system regenerates its own decision policies. A newly generated policy directs no capital until it has proven itself out-of-sample against the incumbent across multiple market regimes. Every failure path falls back to a known-safe baseline.
A self-tuner with a governor
The system tunes its own settings, and a higher-order governor watches it do it. That governor monitors the self-tuner for over-reaction, oscillation, or drift and adjusts its caution accordingly, so the system can neither thrash itself nor quietly freeze.
Learning from the roads not taken
When a risk guardrail blocks a trade, most systems learn nothing from it. CactusQuant replays the market to measure what would have happened, turning the road not taken into a learning signal, behind a firewall that keeps that signal from ever teaching the AI to evade its own safety limits.
Honest by construction, humans in the loop
The system grades itself only on realized results, never optimistic backtests, and it is built so it cannot peek at the future when scoring its own decisions. The safest adjustments apply automatically; anything that moves real risk waits for a person to approve it, and every change is logged and reversible.
It runs on a paper account. That is the point.
To be unambiguous: CactusQuant currently operates on a paper account, and we make no performance or returns claim. We are not here to advertise a trading edge. The innovation being protected is the safety and self-improvement machinery itself, designed so that no single faulty model, mislabeled result, or runaway tuning step should put the account at risk.
The headline is not “an AI that predicts the market.” It is an AI that can safely improve itself, under guardrails engineered so that getting better never means getting reckless. That is the part we believe is worth protecting.
Patented engine, trade-secret know-how
We patented the engine and its guardrails. The proprietary work stays under the hood: our underlying models, parameters, and datasets are not part of the filing and remain protected as trade secrets. That deliberate split, patents on the systems that make the platform unique and trade secrets on the curated data and know-how, is how Osonwanne Group builds defensible technology across its companies.
Patent pending — U.S. Provisional Patent Application No. 64/098,125 (filed June 24, 2026), held by Osonwanne Group LLC.
CactusQuant is one of two patent-pending ventures Osonwanne Group filed this year, alongside CactusComply, our AI-native Arizona tax-compliance platform. You can read more about both filings on our patents page, see the full portfolio under ventures, or learn how we work on the about page.