Executed Markov 2 on TAO but seeing inconsistent results. What is the correct approach to fine tune it?

Before writing a single line of EA code, I spent a long session making sure the foundation is solid. Here's what got done: 1. Risk rules first, in hard numbers I wrote a dedicated risicoregels.md (risk rules) file. Every number in it becomes a fixed constant in the code, not a setting anyone can tweak: - Demo account only, €250 start, EUR/USD only, max 1 open position - Risk per trade: 0.5% of balance below €500, 1% from €500 - If even the minimum lot (0.01) risks too much → no trade, no "just this once" - Daily stop: −2% in a day → no new trades until tomorrow - Max spread 1.5 pips, no new trades 30 min before/after major news, everything closed Friday 16:00 New York time - Loss limit with two safety nets, whichever is stricter: one third of the highest equity is never touched, and half of the highest profit is always kept. At the start that means the account can never drop below €83.33. 2. Thinking in pips, not euros My first instinct was "keep the profit target tiny so I don't give it back." The maths showed the opposite: a €0.50 target against a €1.25 risk needs a 76% win rate just to break even. So instead: - Take profit = 1.5× the stop distance (in pips) - Once a trade is 1× the stop distance in profit, the stop moves to entry +1 pip → that trade can no longer lose - With that setup, a ~44% win rate is enough to break even Small fish swimming with the whales, but letting the current carry you a bit. 🐟🐋 3. Clear roles between me and AI I briefly wanted to give Claude "carte blanche" to build and run everything. We talked it through and settled on this split: Claude researches, proposes (with reasoning) and builds; I decide and approve every step. Hard limits can only be changed by me, in the code. No AI, and no person, gets uncontrolled access to my money. 4. Python ↔ MT5 bridge working - Took a Hyper-V checkpoint of my trading VM first, so there's always a way back - Found out the MetaTrader5 Python package doesn't support Python 3.15 yet (max 3.14) → installed 3.14 next to it and created a dedicated virtual environment just for Kairo - Wrote a read-only connection test script: it connects to MT5, checks the account, balance and symbol data, and never places an order

Im trying too improve it so far its only on paper trades testing strats but i mainly need the bot too improve

First agent: billing drawings & terrain models. Highest score, targets exactly the bottleneck from Day 2, and lower-risk to validate than a trading bot. Success = a correct, ready-to-use quantity takeoff/DWG that you only need to check instead of recalculating from scratch.