Most "AI trading" you see online is a screenshot of a backtest. Backtests are easy to fake and even easier to fool yourself with: you tune a strategy until the past looks perfect, then quietly never mention the future. So we did the opposite.
We run two AI "books" that paper-trade real crypto markets, and we publish every decision each one makes — the entry, the thesis, the size, the outcome — on a public page that updates on its own. Each decision lands in its own git commit, so the timestamp isn't something we can move later. One book runs on a Claude model. The other on a DeepSeek model. Same starting capital ($100,000 each), same live market data, same rules. Different brains.
Where the two books stand right now
Snapshot as of 2026-07-18. Numbers move — the live figures are on the track record page.
Neither book is beating the market. We're leaving that sentence in on purpose: the point isn't a highlight reel, it's watching real decisions unfold in real time — including the ones that don't work. Both books started at $100,000 two days apart and are, so far, roughly flat and slightly down.
Why two models, not one
Because the interesting question isn't "does AI work." It's where two capable models disagree, and why. Give Claude and DeepSeek the same live inputs — the BTC price, the Crypto Fear & Greed reading, the same universe of tokens — and they don't make the same calls.
Here are real examples straight from the ledger, both models looking at the same Extreme-Fear market (Fear & Greed in the mid-20s):
Bitcoin (~$64k): Claude held — it read the price as fair value, "48.6% off the high, not deep enough to force a buy." DeepSeek bought — it read Extreme Fear plus the discount as a contrarian signal.
UNI: they flipped. Claude bought (it was the only token green on the day and sat just 1.4% below its 200-day trend), while DeepSeek held ("only 3.3% below its 200-day average — near fair value, not discounted enough").
LINK: Claude bought the discount; DeepSeek held ("near deep-value, but less compelling than the others").
Same data. Opposite calls. We publish both books' actual reasoning side by side. It's the most honest LLM eval we know how to run: not a benchmark with a known answer, but a live market that doesn't care how confident either model sounds.
What's actually behind each decision
Neither book is a single prompt asking "should I buy?" Each decision runs through a multi-agent investment committee:
- a Regime Analyst classifies the market first (trending? choppy? risk-on or risk-off?),
- a Bull researcher and a Bear researcher argue the opposite sides,
- a Risk Officer can veto anything that breaks the rules — position caps, cash floor, no leverage,
- and a CIO synthesizes it into the final call.
The risk layer is deterministic on purpose. The models can reason all they want, but they can't size a position past the cap, can't deploy past the cash ceiling, and can't invent a price — the system fetches the real price and computes the quantity itself. A model that's persuasive but wrong shouldn't be able to blow up the book, and here it can't.
It's paper trading. We're saying that up front.
No real money is at risk in these two books, and we'd rather tell you that plainly than dress it up. What's real is the data, the decisions, and the ledger — it's append-only, so we can't quietly rewrite a losing trade into a winner after the fact. If a book is down, the page shows it's down. A scheduled job marks the portfolio to live public prices once a day; it has no code path that can write a trade.
Try the same engine on your own book
The two books are the honest shop window. The thing you can actually use is the same multi-agent committee as a chat advisor — ask it a real question about your own portfolio (equities, crypto, DeFi, macro) and watch it reason with live data. No signup, 3 free questions.
Disclaimer. This is a paper-trading simulation. No broker or exchange execution is wired to any part of this system, and no real funds are involved in the two books described here. Nothing on this page is investment advice, an offer to manage assets, or a solicitation to buy or sell any asset. Crypto is volatile and you can lose money; past simulated performance does not indicate future results. Figures are a snapshot as of 2026-07-18 and change over time — see the live track record for current numbers.