Paper trade Bitcoin with a chart-reading LLM agent that writes its own rules

The author, a developer in Wrocław, built Semantic Signal, a paper-trading agent for BTC/USDC on the four-hour timeframe. Each cycle it pulls about 1,000 candles, computes 79 indicators, renders a candlestick chart image, reads a handful of news items from RSS and Reddit feeds, and recalls three similar past trades plus learned rules and post-mortems from a ChromaDB store. A model (DeepSeek by default, with Gemini, OpenRouter and a local server as fallbacks) then returns one decision with a stop, a target and a size. Every few closed trades it writes new rules about what worked and what it got wrong, and those rules fade over a couple of weeks unless evidence keeps them. Code caps each position at 10% of capital and limits it to whitelisted symbols. After ten months and 501 commits the author re-audited every claim in an earlier post against the code.

At the time of writing the public dashboard showed 11 closed trades, a 36.4% win rate, +0.09% P&L and $10,008.54 of paper capital, at about $6 of model cost for the current session. The author calls eleven trades noise, not a track record. The audit retired several earlier claims: the chart is 1080p not 4K, there is no Kelly sizing, six execution guards were really two, and the bot was never free to run because model calls cost money. A hand-written chart pattern detector was deleted after the model read the chart images better on the author's sample.

No real money is involved; a separate executor last touched an exchange testnet in August 2026 and profitability is unproven. The expected-value check is only a prompt instruction, so the model can still propose a trade the math argues against. The author stresses that the 2,400 tests prove the plumbing, not the model's judgement.