Theoriq partnering with Targon to power AI experimentation

Curating risk-managed yield is, underneath, a research problem. Long before capital is deployed, we want to know how markets behave, where they tend to break, and how much of that can be seen coming at all. Work like that runs on sustained, heavy computation, and that research is now being powered by Targon.

We wanted to share some of the experimentation and work that Targon gives us the capability and compute we need to explore. While this is not an exhaustive list of deep dive into the specifics, we wanted to share at a high level: what we set out to test, how we hold every result to a standard fixed in advance, and why none of it would run at this pace without the infrastructure underneath it.

A team on the cutting edge requires technology ready to make it possible. Targon is providing that to Theoriq.

"Targon is proud to partner with Theoriq to assist with their compute needs," Targon Head of Growth Harrison Adams said of the partnership. "Targon remains committed to providing encrypted compute at best-in-class pricing to start-ups helping advance a decentralized world."

What We Are Testing With AI

Some of our current AI research comes down to one question: can a model sense what a market is about to do over a short horizon, and how honestly can it say so?

We separate three things that often get blurred together:

Knowing that a market is likely to move.

Knowing that uncertainty is rising.

Knowing which direction prices will go.

These are different claims, and the last is the hardest. We are openly skeptical it can be reliably predicted at short horizons at all.

So what we are testing is models that forecast the shape of near-term market behavior: the range of likely outcomes rather than one confident guess about direction. The reason is straightforward. Sensing volatility and instability is what disciplined risk management rests on, and it does not require pretending to know things we cannot.

AI running the science, not just the models

The less obvious part is that AI is not only the thing being tested. It is also doing a large share of the testing.

The program is executed by a set of specialized AI research agents, each with a written charter and a narrow mandate: data integrity, target definitions, baselines, architecture, training protocol, evaluation, and one that governs the whole program against a frozen specification. Each is scoped narrowly and checked against the others, so no single step gets to decide its own result.

The one we rely on most is an adversarial agent. Its only job is to refute the rest. It rederives results independently from raw data, audits every assumption, hunts for leakage, re-computes everything from primary outputs rather than trusting any prior report, and attacks each claim before it is allowed to stand. It has caught real defects, including subtle ones that would have produced a false positive, which were then fixed and re-checked until nothing critical remained. Several of our own early findings were retracted this way, and the retractions stay on the record.

That is the part of "using AI" we think actually matters: not a model writing summaries, but a structured, adversarial process holding both the models and the people running them to a standard set in advance.

How we keep ourselves honest

Every claim is written down and frozen before the experiment runs: what counts as success, what the scoring rule is, and what baseline has to be beaten. Every model has to beat simple, well-tuned baselines before it earns any credit.

We run a battery of checks built specifically to catch us fooling ourselves: deliberately shuffled data that a real signal must fail on, planted leaks the harness must detect, placebo signals that must score at chance. And we keep a final evaluation window under lock, sealed so it cannot be read even by accident, that no one and no agent looks at until the approach and the scoring rules are frozen. It is opened exactly once, and whatever it says is the answer.

Negative results are results

Much of what this program establishes is what does not work, and we count that as output, not failure. Short-horizon direction, tested several independent ways, has so far come back exactly as our skepticism predicted. What survives is narrower, and narrower claims are the ones worth having. That is precisely the point.

This is research, not a product claim. The point of it is to learn what is genuinely learnable, and to be precise about what is not.

Where the work runs

Discipline sets the standard. It does not supply the compute. Research like this is bursty and compute-hungry by nature: quiet stretches broken by sudden demand for large blocks of GPUs, on short notice, for training runs and evaluation sweeps. That is the profile Targon is built for.

As a confidential-compute network from Manifold Labs, Targon supplies GPU capacity on demand, with hardware-level guarantees that keep a workload and its data private even from the machines running it. For us, that means standing up the infrastructure a program like this needs, when we need it, without operating a data center to do it, and it means the confidentiality of our research stays intact end to end.

We could not run this work at the pace or the rigor it demands without a compute partner we trust to deliver. Targon is that partner. The discipline is ours to keep. The capacity to test at this scale is theirs to provide.

‍

About Theoriq

Theoriq builds the risk and yield intelligence layer for tokenized markets. It turns tokenized assets into risk-managed yield through multi-asset and infra-agnostic DeFi vaults: curators set the strategy and the risk limits, and AI-assisted systems execute and monitor within them. Its flagship vault, AlphaVault ETH, applies this framework to ETH-native yield, and the Theoriq Gold Vault extends it to tokenized gold, with plans to extend the model to additional real-world assets. Learn more at theoriq.ai.