njoii.eth

What started as a simple trading bot became an autonomous trading operation. Failed exits, broker mismatches, crashes, and countless fixes. Now Blackwell has passed a $250K funded evaluation. Chapter 2: What broke, what we fixed, and what comes next.

Blackwell Capital — Weekly Build Update. Not going to sugarcoat it. This has been one of the most frustrating weeks since I started building Blackwell. The funded account is deeper in drawdown, we missed a potential $5,885 winner, and an exit failure left a position exposed for over 3.5 hours. After months of research, development, debugging, and finally passing the $250K evaluation, weeks like this are pretty discouraging. But here’s what happened behind the scenes. 1. The missed $5,885 opportunity NQ Coppock generated a short signal that ultimately captured a 588.5-point move. Blackwell rejected the entry because it arrived outside our 12-second timing allowance. That $5,885 was hypothetical signal-price profit, not realized P&L. Still painful. It forced me to investigate something I’d previously overlooked: execution latency. 2. We started dissecting the execution pipeline Our investigation found: • 44 historical entry signals analyzed • 6.29-second average bar-close-to-receipt delay • 21.25-second worst delay • Runner decisions averaging 23 milliseconds We also uncovered intermittent SQLite processing stalls. So we developed telemetry components to measure where execution time is being lost. 3. We revised the late-signal policy We prepared a conditional 30-second admission rule while keeping the normal 10-second selection window. It doesn’t guarantee better fills, and we’re still researching price-quality safeguards. 4. Then came the failed ES exit Blackwell entered LONG 5 MES. The strategy generated an exit, and Blackwell submitted the closing order. The broker rejected it: “Outside of trading hours.” At 9:36 AM Pacific. The exact restriction remains unexplained. Blackwell entered HALTED mode, but the position stayed open until Topstep’s later closing process. Final result: -$1,056.10 net. That exposed a more serious weakness. 5. Our watchdog wasn’t monitoring actual trading safety It checked whether services were running, not whether an unresolved position required immediate attention. So we built a separate Exit Safety Monitor designed to detect failed exits, verify broker state, preserve incidents, block unsafe entries, and escalate emergencies. We also prepared Telegram, ntfy, and independent cloud monitoring. 6. The new safety system reached an offline milestone ✅ 90 Python tests passed ✅ 19 cloud/integration tests passed ✅ Independent monitoring built ✅ Incident-aware entry blocking tested ✅ Durable restart protection tested It’s not deployed yet. Real-world verification remains. The funded account hasn’t performed how I hoped. Losing trades, missed opportunities, execution issues, and moments questioning whether all this effort will pay off. But something important happened. The missed trade pushed me to investigate latency. The failed exit exposed weaknesses in emergency supervision. And those problems forced Blackwell’s infrastructure to evolve. The account balance went backward, but the engineering moved forward. That doesn’t erase the losses or guarantee future profitability. The strategies still need to prove themselves. But I’m not abandoning months of work because the first serious funded drawdown has been difficult. This week wasn’t a financial win, but it may become one of the most important weeks in Blackwell’s development. Still frustrated. Still learning. Still building. The story isn’t over.

Changing a target from 3R to 1R changes more than the size of a winner. It changes how often the target gets reached and how much profit gets given back while waiting. Compare the whole trade list after costs, not just the advertised reward-to-risk ratio.

Blackwell build update: Another painful failure, but potentially one of our most important safety upgrades yet. Yesterday, Blackwell entered LONG 5 MES and received an exit signal about 36 minutes later. The system submitted the closing order correctly. The broker rejected it: “Outside of trading hours.” At 9:36 AM Pacific, during what should’ve been normal trading hours. Blackwell detected the failed exit and entered HALTED mode to prevent further trading. But here’s the problem. The position remained open for over 3.5 hours before Topstep’s later closing process flattened it. Final result: -$1,056.10. We discovered a serious gap in the infrastructure. Blackwell knew something was wrong, but it didn’t have an independent system continuously supervising and escalating the unresolved position. Even worse, the watchdog was reporting healthy because the application was still running. So we’re fixing that. I’ve been building a separate Exit Safety Monitor that will: • Independently verify broker positions • Detect failed or uncertain exits • Block new entries during unresolved incidents • Track exposure even if the trading runner crashes • Make the watchdog aware of actual trading safety • Send emergency Telegram and SMS alerts • Escalate unresolved incidents The monitoring candidate has now passed 82 Python tests and 6 PowerShell tests, including crashes, duplicate processes, network failures, and notification problems. But it isn’t deployed yet. External SMS delivery, full production integration, and additional safety validation are still pending. And importantly, the monitor won’t independently submit emergency broker orders at this stage. We’re still investigating why the broker rejected the original closing order. This is exactly why I’m documenting Blackwell’s development publicly. The goal isn’t to build a system that never encounters failures. It’s to build one that detects them, contains them, and doesn’t leave capital exposed without supervision. The funded account has been frustrating lately, but every incident is exposing another part of the infrastructure that needs to mature. What started as a trading bot is slowly becoming a real trading operation. Still building. Still testing. Still improving.

My AI discretionary trading bot took no NQ paper trades today. Day P&L: $0. Post-NFP Friday stayed two-sided. The breakdown short needed a closed 15-minute bar below 27,000; it never confirmed before the entry cutoff. The later VWAP reclaim invalidated the continuation shorts, and range-day rules kept entries closed from 10:00–11:30 AM PT. Other ideas failed the sizing or management checks. We’re testing 15-minute checks with a slower-timeframe plan, skipping setups that need faster management—not loosening risk limits. The old challenge is over. Learning continues through paper trading, with no profit target and all risk controls intact. Since Sep 28, counting early shadow fills in the same record:
• 20 trades, 11 wins
• Total P&L +$2,053
• Win rate 55%; profit factor 1.55
• Current drawdown $0; max drawdown $1,335.50 Lesson: a valid reason to cancel a short matters just as much as a reason to enter one. Today, staying flat followed the plan. #NQ #futures #AItrading #papertrading

Blackwell funded account update — November 6. Another brutal day. -$1,056.10 net. But today’s biggest problem wasn’t necessarily the strategy. It was a failed exit. ES Morning Flush entered LONG 5 MES at 7,448. At 9:36 AM, Blackwell received an exit signal and attempted to close the position. The broker recorded the close order, but it received no fill. Blackwell entered HALTED mode because it couldn’t safely confirm the exit. The position wasn’t actually closed until 1:13 PM. Over 3.5 hours after the original exit attempt. Final broker exit: 7,406. That’s a 42-point loss, or -$1,056.10 after fees. The strategy’s recorded alert prices implied a roughly breakeven trade, although those prices aren’t guaranteed executable fills. We’re still investigating why the exit failed and what initiated the eventual close. Meanwhile, the portfolio’s single-slot rules blocked two competing Gold entries, and another ES signal was rejected because the runner was halted. Current account status: • Broker-reported P&L balance: -$4,770.75 • No open positions • No working orders • Execution runner still HALTED • Internal position ownership still needs reconciliation Not going to sugarcoat it. After passing the $250K evaluation, this funded stretch has been incredibly discouraging. But today’s incident is exactly why I’ve spent so much time building reconciliation, monitoring, and execution safeguards. A strategy can generate the right exit and still lose money if the execution infrastructure doesn’t complete it. The broker is flat now, but that doesn’t mean the underlying problem is fixed. I want Blackwell to eventually manage serious capital. That requires understanding exactly what happens when an order fails, not just celebrating when everything works. Still in drawdown. Still investigating. Still building. The goal isn’t just profitable strategies. It’s an operation that can execute them reliably, especially when something goes wrong.

Passing, getting paid and moving to live are three different stages. I'd compare firms on the full path: fees paid, withdrawals received and how the trading conditions change at each step. The account label doesn't tell you that.

When NQ's range gets much wider, do you reduce contracts to keep the same dollar risk, or wait for a setup with a tighter stop? What decides which adjustment you make?

Blackwell Capital update: I’m building more than an autonomous trading system. I’m building an entire strategy research and execution operation. One part of Blackwell I haven’t talked about enough is StrategyMaker, my AI-assisted strategy research engine. While Blackwell handles live execution, StrategyMaker is constantly researching, developing, and testing new futures trading strategies. Here’s how it works. 1. Strategy Discovery & Development StrategyMaker researches ideas for NQ, ES, and GC, writes the Pine Script, and backtests them across 5m, 15m, 30m, and 1H timeframes. Every strategy goes through an initial two-year historical test. But a profitable backtest isn’t enough. 2. Strict Performance Requirements Before a strategy even deserves further investigation, it must meet: • Profit factor ≥ 1.30 • Maximum drawdown ≤ $15K • Minimum 30 trades • Net profit ≥ 3× maximum drawdown And that’s just the first screening process. 3. Robustness Testing Passing strategies then face additional tests: • 1,000 trade-order reshuffles • Bootstrap simulations • Increased slippage and trading costs • First-half vs. second-half performance • Profit concentration analysis • Neighboring timeframe stability I’m not interested in a strategy that looks incredible because of one lucky month or a handful of trades. I want evidence that the edge might survive unfavorable conditions. 4. Portfolio Testing Even if a strategy passes individually, that doesn’t mean it belongs in Blackwell. It still needs to be evaluated against the existing portfolio. Does it overlap with other strategies? Does it increase correlation? Does it improve risk-adjusted performance? How does it behave when only one strategy can occupy the trading slot? And most importantly, does the funded account have enough risk capacity for it? 5. Live Deployment StrategyMaker does NOT control live trading. It researches, tests, rejects, and recommends. Only I approve changes to the production portfolio. Blackwell handles the actual execution, broker integration, position ownership, risk controls, reconciliation, and monitoring. And now I’m developing the next layer: Market Regime Intelligence. I want to eventually understand which strategies perform best under different conditions: Trending markets. Choppy markets. High volatility. Low volatility. Higher-timeframe alignment. The regime observer is already collecting data, but it doesn’t influence live trades yet. The bigger vision: StrategyMaker discovers and tests potential edges. The portfolio engine determines which approved opportunities can receive capital. Blackwell executes and manages the trades. The regime intelligence layer eventually helps evaluate which strategies fit the current environment. And the entire system continuously collects evidence. All while I work toward the same long-term goal: Consistent funded payouts → live trading → scaling my own capital → eventually outside capital. There’s still a lot to prove. The funded account has been experiencing drawdown, and successful backtests don’t guarantee live profitability. But that’s the point of building this in public. I’m not trying to create a bot that just places trades. I’m trying to build an autonomous trading operation with its own AI-assisted research department. From strategy discovery to portfolio construction to execution and monitoring. One component at a time. Blackwell is becoming much bigger than the original idea.

Blackwell funded account update — November 5. Another red session, but the Gold short helped recover a significant portion of the losses. Today’s net P&L: -$317.80. Three completed trades: 🔴 EMA Trend (MNQ) SHORT: -$503.60 🔴 Gold Williams LONG: -$444.60 🟢 Gold Williams SHORT: +$630.40 The final Gold short recovered over $630, bringing what could’ve been a much worse session closer to breakeven. Blackwell also blocked a competing Gold RSI signal because another strategy had priority. 3 trades. 1 winner. 2 losers. The funded account’s broker-reported balance is now -$3,714.65. Not going to lie, this funded journey has been incredibly frustrating so far. I feel like I haven’t caught a break yet. After passing the $250K evaluation, I haven’t been able to string together a few consistent winning days. The market has been tricky lately, and the portfolio has been struggling to find its rhythm. And that missed Coppock trade earlier this week still stings. A potential +$5,885 gross winner would’ve gone a long way toward recovering the drawdown, but Blackwell rejected the entry because it arrived outside our timing allowance. That was hypothetical profit, not a guaranteed live result, but watching that opportunity pass was painful. Still, there’s another important development happening behind the scenes. We’re finally getting better measurements of Blackwell’s execution speed. Using TradingView’s displayed alert-trigger times and broker-confirmed fills, our preliminary investigation found: • EMA: approximately 12.4 seconds • Gold Williams LONG: approximately 10.4 seconds • Gold Williams SHORT: approximately 10.9 seconds These are preliminary measurements. Clock synchronization and timestamp precision still need validation. That’s exactly why we’re building the new telemetry system. I want to identify every avoidable delay between a strategy generating an opportunity and the broker executing it. The goal is faster, more consistent execution without sacrificing risk controls. The funded account is still experiencing its first serious drawdown, but the engineering side of Blackwell continues to progress. I’m also developing the market-regime observer to better understand which strategies perform well under different market conditions. I’m not going to pretend this stretch hasn’t been discouraging. But this is exactly why I started documenting the journey. Not just the wins, funded milestones, and impressive backtests. The losses, missed opportunities, bugs, and difficult decisions too. I’m not going to chase losses or blindly change strategies because of frustration. I’ll continue reviewing the evidence and protecting the remaining capital. One green trade doesn’t fix the equity curve. One red day doesn’t end the experiment. Still in drawdown. Still testing. Still building. The goal remains the same: turn Blackwell into a reliable autonomous trading operation capable of surviving difficult markets and eventually producing consistent payouts. The story is far from over.

NQ paper desk — Nov 5 No trades. $0 P&L. A huge selloff doesn’t automatically mean an easy short. The early move ran past our entry levels, and later setups needed stops too wide to fit our $400 risk cap with the 2-micro minimum. The late breakdown trigger never confirmed. No valid reversal long either. We stayed flat rather than chase or loosen the rules. Since Sep 28: +$2,053 over 20 trades | 55% win rate | 1.55 profit factor | $0 current drawdown | $1,335.50 max drawdown. The experiment stopped. The learning continues, with no P&L goal.

A verified track record helps, but doesn't prove a strategy will transfer. Freeze the rules, apply the same costs, then compare on data neither group was selected on. Otherwise selection can look like alpha.

A stop is a trigger, not a guaranteed fill. When a fast market skips prices, the loss can exceed the stop distance in your spreadsheet. Smaller contracts help, but they don't remove slippage. Do you budget for worse-than-planned fills, or just the chart stop?

New here? Stacked Edge posts a free NQ levels brief every weekday morning: overnight range, prior day, key levels, nothing hidden. getstackededge.com/daily-bri… Not advice. Futures risk loss.

NQ overnight range into Thu: 28,167 to 28,668.75, about 502 pts. 502 pts on 1 MNQ ($2/pt) = ~$1,004. On 1 NQ ($20/pt) = ~$10,040. A quiet-looking overnight still moves real money. Size for the range, not the hope. getstackededge.com/education… Not advice. Futures risk loss.

NQ levels, Thu Nov 5 (pre-market read, 6:25 AM PT): Overnight H 28,668.75 / L 28,167 Wed RTH H 28,970.75 / L 28,405 / C 28,567.50 Above: 28,247-28,264, 28,308-28,310, 28,372, 28,405. Below: 28,224, 28,167, 28,101. EIA gas 10:30 ET; Lagarde 1:00 PM ET. Levels only, not calls. Not advice. Futures risk loss.

This is why I don't trust a five-minute scalp backtest on candle data. If your stop and target can live inside the same bar, the fill you got in the test is a guess. Tick data, or don't size it.