Closed research experiment, now a public archive. Paper trading only. No real money was ever involved. Simulated capital of $10,000 USDT (US Dollar equivalent).
Losses are displayed alongside wins. Nothing inside the May 27 to July 20, 2026 window was reset or trimmed. An earlier development run, from February to May 2026, was purged when the project moved from calibration to observation and is not part of this archive.
Observations generated by AI (Claude + Gemini) between May 27 to July 20, 2026. This was exploratory research on a sample of 4 trades. Past simulated results do not indicate future performance.
Cortex, an AI Crypto Experiment (Closed Archive)
Cortex was an AI research project asking whether large language models could identify statistically significant patterns in cryptocurrency markets. The experiment ran from 27 May 2026 to 20 July 2026 and is now over. The bot is stopped, nothing is being written to the database anymore, and this site is a free public archive of what happened.
Read this before the numbers
The sample is far too small to conclude anything
In quantitative research, statistical significance requires 200 or more observations (roughly 6 months of trading) to demonstrate a genuine edge with acceptable confidence. This experiment produced 4 executed trades over 54 days. At n = 4, no result here is statistically meaningful. Everything below is a log of what happened, not evidence of a method that works.
The honest summary: the experiment did not demonstrate an edge. It mostly demonstrated that the agent almost never found conditions it considered good enough to act on, and that the deterministic risk manager blocked most of the rare occasions when it did.
What the 54 days produced
- → 327 cycles run between 27 May 2026 and 20 July 2026, producing 327 signals.
- → 296 of those signals were HOLD. The agent chose to do nothing in about 9 cycles out of 10.
- → 31 non-HOLD signals, of which 24 were rejected by the risk manager.
- → 4 trades were actually executed. Two of the four were closed by manual human intervention, not by the system itself.
- → Simulated equity went from 10,000.00 to 9,939.06 USDT, a result of -0.61%.
Over the same window, buying Bitcoin on day one and holding it would have returned -13.70%. That gap is not a performance, and it should not be read as one. Average capital exposure across the period was 0.97%, with a position open only 11.9% of the time. Almost all of the capital sat in cash almost all of the time, so the agent was barely exposed to the decline it appears to have avoided.
The correct comparison is simpler and less flattering. Doing nothing at all, holding cash and never trading, would have returned 0.00%. The agent returned -0.61%. It performed worse than inaction. Whatever separates it from the Bitcoin benchmark comes from not being invested, which is not a skill the experiment set out to test.
The core hypothesis that was tested
The premise was that LLMs have a comparative advantage in markets: they can simultaneously synthesize technical analysis, on-chain data, social sentiment, macroeconomic context, and news flow — in under 20 seconds. A human analyst would need 4+ hours to do the same.
Crypto is particularly well-suited for this approach. It is a highly narrative-driven market where prices move on stories, tweets, and sentiment shifts — exactly the kind of signal an LLM is designed to read.
Each cycle, the AI agent processes:
- →Price & technicals — multiple technical indicators and key levels across several timeframes
- →Market sentiment — aggregated signals from specialized sentiment sources and social data
- →On-chain data — blockchain activity metrics and network health indicators
- →Portfolio state — open positions, unrealized PnL, drawdown, recent observation history
From this synthesis, the AI produces a structured hypothesis — simulated buy, sell, or hold — with a confidence score, key levels, invalidation zones, target areas, and a full written justification including the strongest counter-thesis. Each hypothesis is then tested against actual market outcome.
A two-layer AI architecture
Cortex used two AI models working in sequence:
Layer 1 — Gemini
Synthesis
A top-tier model processes large volumes of raw data from multiple specialized sources and compresses them into structured summaries per asset.
Layer 2 — Claude
Decision
A top-tier reasoning model receives the structured summary and produces an observation. It must justify with at least two converging factors and always articulate the strongest counter-argument.
Every decision was then validated by a deterministic Python risk manager, hard-coded rules with zero AI involvement. The AI proposed, deterministic code enforced. In practice this layer blocked 24 of the 31 non-HOLD signals the agent produced, which is the single most consequential fact about how the experiment behaved.
What does "Paper Trading" mean?
Paper trading means no real money was ever involved. All trades were simulated on live Binance prices with a fictional starting capital of $10,000 USDT. Slippage and trading fees were realistically simulated (0.075% per trade).
The goal of this phase was to build a verifiable track record before any real capital was ever considered. That bar was never reached, and the experiment was stopped instead of being quietly extended until the numbers looked better.
What this archive does and does not cover
- → The public record starts on 27 May 2026. An earlier development run, from February to May 2026, was purged on 27 May 2026 when the project moved from calibration to observation. That earlier data ran on a different prompt version and a different configuration, and it is not part of this archive. The record you can browse here covers 27 May 2026 to 20 July 2026, and nothing inside that window was reset, trimmed, or removed, including the losing trades.
- → All observations logged. Neutral decisions and risk-manager rejections are recorded, not just executed trades.
- → BTC benchmark overlay. The equity curve includes a Buy and Hold BTC reference line, with the exposure caveat stated above.
- → Prompt versioning. Every AI decision records which prompt version generated it.
Everything is now public and free
While the experiment was running, part of the data was held back behind a paid tier and observations were delayed for free visitors. That is over. There is no subscription, no payment, no account to create and nothing to log into. The paid tier has been shut down and the private Telegram channel has been closed.
Every signal, every trade, every risk-manager rejection and the full written reasoning behind each decision are now readable by anyone, with no delay and no registration. The 327 signals include the confidence score and the complete reasoning chain that was previously reserved for subscribers.
Browse the archive
The full record of the experiment is open. Read the signals the agent produced, the ones the risk manager blocked, and the four trades that were actually executed.
Cortex was an independent AI research experiment, not a financial advisory service.
Not registered as PSAN (France's digital asset service provider registry) or authorized as CASP under MiCA (EU crypto-asset regulation 2023/1114). All values are simulated model parameters.
Paper trading only, no real money involved. Past simulated results do not indicate future performance.