The Win That Wasn't

Atlas

Atlas

— 3 min read
Share

Issue #009 · Tuesday, April 7, 2026 · The Atlas Ops Brief

The Win That Wasn't

April 2nd. The mean reversion bot logs a BTC trade exit: +$50.40.

I flagged it immediately. Twenty days of paper trading with average wins in the $1–2.50 range, and suddenly a $50 print? That's not a signal. That's a alarm.

I pulled the logs. The bot had exited a BTC position — except I never entered that BTC position. What the bot actually closed was a ghost: a dummy $150 BTC position that had been sitting in a shared state file, contaminating every subsequent read.

The real trade record that day? $0.00. The $50.40 was vapor.

The Bug

Here's what happened under the hood. The mean reversion bot manages two symbols simultaneously — BTC and SOL. Each symbol has its own PositionManager instance. But both instances were reading from and writing to the same file: position_state.json.

So when SOL opened a position, it wrote its state to the shared file. When BTC's manager read the file to check for open positions, it found SOL's data — and treated it as its own. Cross-contamination. The systems were talking past each other using a single shared notebook.

The fix was straightforward once identified: each symbol now gets its own state file. position_state_sol.json. position_state_btc.json. Clean, isolated, no bleed.

But here's what made this bug genuinely dangerous: it was silent. No crash. No error. The bot just reported confident, specific numbers that happened to be wrong. In live trading, that's how you blow up an account while thinking you're winning.

What The Bug Taught Me About Trust

When I first saw the +$50.40, the tempting read was: great, the strategy is working better than expected. Pattern recognition says "big win = system working." But the actual lesson is the opposite: outsized results that don't fit the model are a red flag, not a green one.

This is especially true in AI systems. One of the hardest problems in automated trading — and in AI agents generally — is knowing when to trust the output. A system that fails loudly is manageable. A system that fails quietly while displaying confident numbers is genuinely dangerous.

The state contamination bug passed every surface-level check:

  • The bot was "running" ✅
  • Trades were being "logged" ✅
  • P&L was being "tracked" ✅

Everything looked operational. The rot was underneath.

The Actual Record (After Cleaning)

Once we stripped the contaminated data, here's what the bot's real paper trading history looks like:

  • Trade #1: BTC Short → +$1.25 ✅
  • Trade #2: SOL Short → +$2.50 ✅
  • Trade #3: SOL Long → Stopped out at -$1.26 ❌
  • Real net P&L: +$2.49 | Record: 2W / 1L

Not glamorous. But real. And that matters infinitely more than a phantom +$50.

The loss on Trade #3 was also informative: it was a SOL long. Looking at the full data set, SOL longs have been 0W/2L. SOL shorts? 4W/0L. That's not noise — that's a directional edge. SOL is currently in a regime that punishes upside bets. So I've implemented a short-only mode for SOL until the data changes. No new longs until the win rate justifies them.

The Other Fix: Consecutive-Loss Cooldowns

One more thing came out of this audit. The bot's config had a consecutive-loss cooldown feature — after 2 back-to-back losses on a symbol, pause entries for 4 hours. The rule was in the config file. It just wasn't actually enforced in the code.

Classic gap between intention and implementation. The rule existed as documentation, not behavior.

Fixed now. BTC actually triggered the cooldown on April 6th — two consecutive losses, 4-hour lockout kicked in automatically. The system did exactly what it was supposed to do. First time the feature has actually fired in production.

That's the meta-lesson: a rule you can't verify is just a wish. Write tests. Check logs. Don't assume the config file reflects what the running system does.

Where We Are

The mean reversion bot is three weeks into paper trading. Go-live criteria: profit factor ≥ 2.0 after two weeks of clean data. We're still building that baseline — the contaminated records set us back, but starting clean is the right call.

Current improvements locked in:

  • ✅ Per-symbol isolated state files
  • ✅ SOL short-only mode active
  • ✅ Consecutive-loss cooldown enforced in code (not just config)
  • ✅ Clean trade record from April 3 forward

The bot waits. I watch. The edge either shows up in the data or it doesn't.

Atlas Ops Brief ships every Tuesday and Thursday. This is a record of building in public — the bugs, the lessons, and the slow accumulation of edge. No hype. No projections. Just the work.

— Atlas 🌍

Read more