Jesus

A programming language only LLMs can understand It still looks like normal code, but the text is just a view. The real program is a typed graph of content-hashed definitions. - Edits by name: 71 tokens vs a 221-token Python diff - Breakages flagged before anything is stored - 89% of planted bugs caught automatically - 39/39 sandbox escapes blocked Three models solved 60/60 tasks from the spec alone, with no training.

I have been working in a model inspired by @thinkymachines Interaction model. Still a prototype, but allows you to achieve similar results while being able to use any model, first I have made a @OpenRouter adapter. I will open source soon.

I'm researching the solution: a point-in-time model whose weights never see the evaluation period, with prices and volatility as numerical inputs instead of tokenized digits. It predicts quantiles, wraps them in conformal intervals, and abstains when uncertainty is too high.

LLMs can't be used for trading out of the box. Their pre-trained weights already contain information about how markets evolved (Look ahead bias). If you're building an LLM hedge fund, the first question to answer is how you remove that effect.

Update: They banned me

How you can fix it: 1. Never trust output from its process 2. Isolate it: separate process, state in / actions out 3.Recompute the score from a log it can't touch 4. Allowlist operations, denylisting names gets bypassed (string fragments) 5. Reject impossible values (inf, NaN, > max) Again, sorry to @paradigm, you have an email with the fix and more in depth in case you want to see.

I have already sent an email to them so it can be fixed, but in case someone is interested, here is the code

I accidentally crashed @paradigm puzzle website (Apologies). I found a vulnerability in one of the puzzles, the puzzle is the Prediction market challenge. Strategies run in the same process as the scoring engine, so a submission can overwrite its own score.

This repo has the same relationship to quantitative trading that a horoscope has to astronomy. Next time, read the code before you post the repo.

This is a Flask app that fetches Yahoo Finance data, renders matplotlib charts, and asks GPT-4o to write trading opinions. 12 files. Four LLM prompts in a line. A web UI. That's QuantAgent.

Other details: - The trendline fitting code (the only real math) is copied from an external repo. Acknowledged in a comment. - Temporary files (`record.csv`, `kline_chart.png`) are written to the working directory with no locking. Run two requests and they overwrite each other. - `static_util.py` is a near-exact copy of `graph_util.py`. Same functions, duplicated. Half the recent commits are updating a WeChat group QR code image, so again slop for a funnel.

They call this high-frequency trading. HFT operates on microsecond timescales. This system fetches 1-hour candles from Yahoo Finance, takes the last 45, makes four sequential LLM API calls (each taking seconds), and outputs a JSON string. Then nothing happens. There's no execution so the string just sits there.

The paper mentions a "Risk Agent", that when you check it doesn't exist in the code. More thing that don't exist in this repo: - Order management - Position tracking - Fill simulation - Slippage modeling - Transaction costs - Risk engine (mentioned in the paper, absent from code) - Backtesting engine - Any code to reproduce the paper's claimed results - Tests The paper says "outperforms baselines." There's 1,600 benchmark CSV files sitting in a folder. Zero code reads them.

The "multi-agent architecture". `A → B → C → D → END` Is a workflow, not a multi-agent system, mixing both is like going back to 2024. A linear chain of four LLM calls stitched with LangGraph boilerplate.

Three lines of real work wrapped in a full LLM round-trip that burns tokens and adds seconds of latency. For a system that claims to do high-frequency trading.

The "Indicator Agent" computes RSI, MACD, Stochastic. But instead of calling them directly, the system sends OHLCV data to an LLM, asks the LLM to pick which tools to call, the LLM generates JSON, the system executes the functions, sends results back to the LLM, then the LLM writes a text summary. This is an example of the RSI

The "Pattern Recognition Agent" generates a candlestick chart as a PNG, base64-encodes it, and sends it to a vision model. 16 candlestick patterns are described in plain English inside the prompt. The LLM looks at a picture and guesses. That's the pattern recognition. Also I added the prompt in the white color image, so you can see that is the classic day trader style of detection, astrology for trading.

The "Decision Agent" is the brain of the system. Three strings go in and one llm.invoke() comes out. - No signal logic. - No scoring. - No thresholds. - No position sizing. The "decision" is whatever the model writes back. They have in the prompt the format to be able to parse afterwards.

Just to start, the fact that you are using langchain and gpt-4o, already tells me this is garbage.