Applied AI market research
Genesis is an open research project studying how AI systems interpret live crypto and equity markets, and whether that interpretation holds up over years rather than headlines.
The project is run in public and in writing. Every assessment the model produces is logged before outcomes are known, so the eventual verdict, favourable or not, rests on a record that cannot be edited after the fact. The full method, including how the project's wallets are used to track progress, is described in the documentation.
Why this study exists
AI systems now write passable market commentary on demand. What remains unknown is whether that fluency translates into durable judgement: the ability to stay roughly right, at roughly the right confidence, across very different market conditions, for long enough to be useful to a real investor. The distance between sounding informed and being reliably useful is exactly what Genesis is built to measure.
Most attempts to answer this question share the same weaknesses. Results are reported over short, self-selected windows. Backtests quietly absorb information that was not available at the time. Failures are abandoned rather than published. Genesis is structured around avoiding all three: fixed rules set in advance, judgements logged before outcomes, and a commitment to publish negative results with the same prominence as positive ones.
The study spans both crypto and equities because the two markets stress different weaknesses. Crypto markets run continuously, move violently, and punish slow reasoning. Equity markets are slower, richer in structured information, and punish overconfidence in narratives. A system that only works in one of them is interesting. A system that works in neither is a clean negative result. Both outcomes are worth publishing.
Research approach
Genesis is designed so that a negative result is as publishable as a positive one. These four principles are what make that possible, and each one exists to close off a specific way that studies of this kind usually fool themselves.
Genesis begins read-only. The model interprets live markets and records intent in writing long before any capital is involved. Execution is introduced only when the quality of the reasoning behind it can be measured, compared and reviewed. This ordering is deliberate: most failures of automated trading systems are failures of reasoning that were never inspected, because the system was allowed to act before anyone could read what it thought.
Every judgement the system produces is written down with the context it was made in: the inputs available at the time, the stated confidence, and the conditions that would prove the view wrong. Nothing is edited after the fact. When results are reviewed later, they are reviewed against this record rather than against a reconstruction of what the model probably believed, which removes the hindsight bias that makes most post-hoc analysis of trading systems worthless.
The question is not whether a model can trade well for a day or a week. Almost any system can look competent over a short window, in a favourable regime, with a sympathetic benchmark. The real question is whether the system remains useful across quarters and across regimes: through trending markets and flat ones, through calm and through stress. Genesis is designed to run long enough for that distinction to become visible.
The target outcome is a durable assistant for a human decision maker, not an autonomous oracle. That means the system is judged on framing, monitoring and preparation: whether it notices what matters, whether it says so early and clearly, and whether a person working with it makes better decisions than a person working without it. A system that produces returns but cannot explain itself fails this test, because its operator cannot know when to stop trusting it.
Programme
The project advances only when the current phase has produced enough written, reviewable evidence to justify the next step. No phase has a fixed end date; each ends when its record is complete enough to judge. This keeps the study slow by design and makes it impossible to rush toward live capital on the strength of a good month.
Phase 01
Continuous read-only ingestion of crypto and equity market data. The model produces written assessments of what it believes is happening and why, with no capital at risk. This phase builds the baseline record that every later phase is judged against.
Phase 02
Assessments become explicit, timestamped hypothetical positions. Each entry states what the model would do, at what size, and what would invalidate the idea. Accuracy, consistency and reasoning quality are then scored against outcomes, producing a track record that costs nothing to obtain but is expensive to fake.
Phase 03
Small, bounded allocations across the project's Solana, Bitcoin and Ethereum wallets, each with hard position limits, per-period loss limits and a full audit trail. Live results are compared directly against the paper record to measure what market friction does to the model's stated intent.
Phase 04
A structured review of whether the system actually improves a human operator's decisions over a multi-quarter horizon, and a written account of where it helps, where it misleads, and where it should not be trusted at all.
Progress through the phases, along with the wallet-level record once live allocations begin, is documented in the project documentation. The documentation is the canonical reference for how the study is run and how its results should be read.
