NERVE protocol keeps GPT-6 Astra behind a six-node trading bot pipeline
An open-source trading bot architecture where 97% of tokens are rejected by a deterministic reflex layer, so GPT-6 Astra is called on under 3% and never sees private keys, addresses or transaction bytes.
this repo is fucking insane and should not be public NERVE Protocol - six nodes, one impulse, cheap reflexes kill a rug before GPT is ever called repo: https://github.com/h100envy/nerve your trading bot calls GPT on every token it sees NERVE calls GPT on less than 3% of them the rest die in reflexes: deterministic checks that cost zero, run in milliseconds, and cannot be argued with the biological metaphor is literal: > an impulse is one Pydantic message carrying facts through six specialized nodes > reflexes fire before any expensive or side-effecting call > the spine routes the impulse and short-circuits on REJECT > every transition is persisted to SQLite, recoverable after a restart the six nodes: > scanner watches Uniswap V3 pools on Robinhood Chain > reflexes check kill switch, daily loss, gas cap, liquidity, slippage, duplicate positions > analyst calls GPT-6 Astra with normalized metrics only > risk sizes the position in pure code > executor signs and sends exactly once > monitor watches quotes and creates exit impulses the model boundary is the whole point GPT receives no private key, no RPC client, no wallet address, no transaction bytes, no gas limit, no position size it returns PASS or REJECT, a confidence score, a thesis and risk flags that is all it is allowed to do a failed model call rejects the impulse, a model that fails is a model that says no reflexes live in code not in prompts, a prompt can be argued with, code cannot ships paper-first, no RPC connection, no key import, no OpenAI call in paper mode the first production adapter targets Uniswap V3 on Robinhood Chain, source interface accepts Base and Solana without changing the protocol
This post quotes another post: x.com/h100envy/status/2100248695552049534