Trading System ● Live Monitor

One paper account ($1,000), two sleeves — built to mirror the eventual real setup of a single brokerage account trading whatever it wants. The STOCK SLEEVE ($900) is the Active book: a diversified, defensive basket of trend-qualified stocks and ETFs — it rides whatever is in an uptrend, sits in cash on whatever isn't, and rebalances monthly (trend-following, because that's what has the strongest long-run evidence). The CRYPTO SLEEVE ($100) trades BTC/SOL/XRP around the clock: it watches hourly but only trades when the 4-hour trend genuinely flips (faster trading was tested and costs eat it); right now it's long BTC, XRP. An AI reasoning layer reviews live conditions every 30 minutes and judges events in context; an autopilot promotes or benches whole strategies based only on live results.

Best recent performer: TSM standalone (risk-adjusted score 1.8); weakest: rl bandit (-0.47). The consistent pattern: defensive trend-following and adaptive blends work; raw momentum without a regime filter doesn't. Important honesty: these rankings are still mostly HISTORY. The true out-of-sample scorecard started accumulating on Aug 1 and is still empty; live results, not history, will decide who keeps capital.

Thesis check: the system profiles every market it watches — of 73 securities, 16 currently behave momentum-driven and 38 mean-reverting, and strategies are weighted to fit each one rather than forced onto all. The core bet — that no single strategy survives regime change, so the edge is in adapting — is so far SUPPORTED by the data: every static strategy we tested decays somewhere in history, while the adaptive blend holds up. Alpha-decay watch: raw momentum's recent score is far below its long-run average (crowding), which is why the book leans defensive trend instead.

Last month the champion book returned +0.3% vs the S&P's -1.0% — it outperformed. Why: the book is deliberately defensive (big dollar-index position, no crypto longs, trend-qualified equities only), so it lags in sharp rallies and protects in selloffs — that's the design, not an accident. Its whole 25-month simulated run: +31.39% with a worst drawdown of -4.59%, i.e. it earns by not losing.

The AI brain is NOT being fine-tuned yet — deliberately. A training set builds itself daily from the system's own decisions and their real outcomes (6,237 examples so far), but an anti-overfitting gate blocks training until the data is diverse and balanced (current blocker: label imbalance (minority class 0.8% < 15%)). Recommendation: keep accruing through varied market conditions; once the gate opens, run a light LoRA fine-tune of the open-weight model on outcome-labeled decisions (teaching it when to veto and downsize, not to predict prices), then A/B it against the untuned brain on the forward leaderboard before it touches allocations. Warranted right now: no.

Exit triggers in force: the crypto book exits within hours when the 4h trend or its regime envelope breaks, and a kill-switch flattens everything at -25% account drawdown (not triggered). The monthly book exits any holding whose long-term trend breaks — but only at month-end rebalance; single positions are small (~3-8% each) and a market-stress scaler halves the whole book after a sharp broad selloff, which bounds mid-month damage from a single blowup (e.g. earnings). Position watch (sent to the brain for judgment — a break can mean deterioration, a shakeout, or a buying catalyst; nothing tactical is auto-sold on a single fact): AGG has broken below its long-term trend.

Path to real money: 2 of 6 gates passed. The missing gates all need LIVE time — months of real forward record with adequate risk-adjusted return and controlled drawdown, plus a month of clean execution on the $100 account. No shortcuts: when all gates go green the dashboard flags it, and going live is still a deliberate human decision starting tiny.