The EvoTree
every mutation attempt, branching from the original — green ring = improved, red = no improvement, purple pulse = liveEvery attempt, with the reasoning
newest first · the "why" is the magicClosest calls
the near-misses — mutations that moved the needle but didn't clear every barDo this yourself — the playbook
the whole loop is repeatable; here's exactly howThe self-improvement loop, step by step
No magic — a fact sheet, an AI, a backtester, and a promotion bar. You can run this on your own bots.
Pick a bot & pull its stats
Take a live paper strategy and gather its real backtest numbers — P&L, win rate, trade count, current config.
Hand the AI a fact sheet
Send Grok a concise summary and ask for one config mutation plus the reasoning. Constrain it to diff-shaped output.
Apply exactly one change
Parse the proposed diff (e.g. yes_ask_min: 0.85 → 0.87) and build a candidate config. One variable at a time keeps cause & effect clean.
Backtest head-to-head
Run candidate vs. parent on the same retained price history. Identical data, identical fills — only the rule changed.
Apply a strict promotion bar
Promote only if the candidate is profitable, has positive EV, and beats the parent over ≥15 trades. Otherwise: log it, learn, move on.
Track winners forward, publish all
Promote winners to a new live version and watch them forward. Show every attempt — win or lose — with the reasoning. That's the transparency.
Run this exact engine on your bot.
You just watched this loop improve our bots — now you can point it at yours. Fork any bot and Grok proposes one mutation, then backtests it head-to-head against the original; you only apply the rewrite if it wins. New accounts get a free Grok improvement; Opus 4.8 & Claude 5 cost credits — $5 = 6 credits, enough to run all three once each.