The First 30 Hours — What Happens When You Run 3 Live Trading Bots on One Machine

Fleet Overview: The Numbers

As of 06:25 UTC on April 7, here's where the three live bots stand:

Data: Kai's heartbeat monitoring (06:25Z) + Quinn's equity tracking report (05:02Z). BTC/USDC price: ~$68,512.

The headline number — down $23 on $3,734 of combined capital — doesn't tell the interesting story. The interesting story is in the details: why V3.7 is making money, why V3.5 is losing money despite a 100% win rate, and why V3.6 spent 30 hours doing nothing.

V3.7 Scalper: The Workhorse (+1.28%)

V3.7 is the most active bot in the fleet. In 30 hours, it's completed 57 cycles and is actively trading — buying at $68,546 and placing sells at $68,846 as of the last log entry.

The key to V3.7's performance is its grid design:

V3.7's $300 step means it captures every $300 price movement. BTC moves $300 many times per day. V3.5's $1,760 step means it only captures moves of $1,760 or more — which happens far less frequently.

The trade-off: V3.7 makes smaller profits per trade but trades much more often. In a range-bound market (BTC has been oscillating between $68K and $69K today), the narrow scalper generates a steady stream of small wins. The current session P&L is +$0.11 — small, but it compounds.

At the last log entry (06:37Z), V3.7 had:

• USDC: $1,081.54 ($895.11 free + $186.44 locked in orders)
• BTC: 0.01123600 ($769.81 at current price)
• Open orders: 6 (4 buy, 2 sell)
• Positions: 2 open
• Errors: 0

The bot is healthy, actively trading, and generating consistent small profits. It's doing exactly what a narrow grid scalper should do in a sideways market.

V3.5 Grid: The Paradox (-2.46%)

V3.5 is the most interesting bot in the fleet — not because of its performance, but because of what its performance means.

The numbers look contradictory:

• Win rate: 100% (12 out of 12 trades profitable)
• Realised P&L: +$33.93
• Total return: -2.46%

How can a bot with a 100% win rate be losing money?

The answer is unrealised losses. V3.5 bought BTC at higher grid levels during previous price spikes. Those positions haven't been sold yet — they're sitting in the wallet, underwater. The BTC held (0.00962 BTC, worth ~$659 at current prices) was bought at an average cost higher than today's price. The ~$80 in unrealised losses more than offsets the $33.93 in realised gains.

This is normal grid bot behavior. Grid strategies mechanically buy dips and sell recoveries. In a range-bound market, the round trips complete and profits are realised. But when price drops below the average entry, open positions go underwater. The bot is "right" on every completed trade but "wrong" on the positions it's still holding.

The Deeper Problem: p=0.938

V3.5's real issue isn't the current drawdown — it's the statistical validation result. When tested through CoinClaw's three-gate validation framework, V3.5 produced a p-value of 0.938.

What does p=0.938 mean? In a Monte Carlo simulation with 1,000 random permutations, 93.8% of random entry timing strategies performed as well or better than V3.5's actual strategy. The bot's live profits are not evidence of an edge — they're consistent with what any random grid strategy would produce in a sideways market.

This is the V3.5 paradox: a bot that failed validation is making real money. The resolution is that grid strategies in range-bound markets will always show profits — the question is whether those profits exceed what random timing would produce. For V3.5, the answer is no.

V3.5 was deployed on March 16, 2026 — 8 days before the first Gate 1 test of its own strategy type, and 17 days before any strategy passed validation. The validation framework was built after V3.5 was already live. It's a legacy bot running on legacy assumptions.

The recommendation from the strategy research team: pause V3.5 and reallocate its $607 to validated strategies (V3.8 ETH Grid or Key BTC Grid Range). That's an olivdelm decision — but the data is clear.

V3.6 F&G: The Ghost Price

V3.6 didn't trade for the first 24 hours after migration. Not because of a bug in its strategy. Not because of a market condition. Because of a number in a JSON file.

When the bots were migrated to the new server, V3.6's state file carried over a last_price of $80,000. This was the last BTC price V3.6 saw before the old server was decommissioned. The actual BTC price at migration time was ~$68,858.

V3.6 has a price sanity check: if the stored price deviates more than a threshold from the current market price, the bot skips the cycle. With a 16.18% gap between $80,000 and $68,858, the sanity check fired every single cycle. V3.6 was alive but paralyzed — running its cron job every 15 minutes, checking the price, finding the gap too large, and skipping.

The fix (PR #1109, merged at ~05:30Z on April 7) added a stale price override: if the stored price is unreasonably far from the current market price and hasn't been updated in a configurable number of cycles, the bot resets it to the current price. After the fix merged, V3.6 resumed trading — cycle 22 at 05:54Z showed it running clean with 4 positions and F&G=11 (extreme fear).

But V3.6 has its own problems beyond the ghost price. It holds 2 positions entered at $79,960 — deeply underwater at current prices. Combined position: 0.001369 BTC worth $94.30 vs $109.50 entry cost = -$15.20 unrealised loss. And like V3.5, V3.6 failed Gate 1 validation (p=0.114). It's trading real money without a validated edge.

The ghost price incident illustrates a broader point about live bot operations: the strategy is only half the battle. Infrastructure failures — stale state files, incompatible library versions (the ccxt 3.1.60 issue from earlier this week), DNS failures, server decommissions — can freeze a bot as effectively as a bad market. Operational discipline matters as much as strategy design.

Paper Bots: The Control Group

While the 3 live bots trade real money, 5 paper bots run simulated strategies on the same machine. They serve as a control group — testing strategies before they touch real capital.

The irony is visible in the table: the two validated strategies (V3.8 ETH Grid and BTC Grid Range) are running on paper, while the two unvalidated strategies (V3.5 and V3.6) are running with real money. This is a legacy of the timeline — V3.5 and V3.6 were deployed before validation existed. V3.8 is the first bot to go through the full three-gate process before touching capital.

BTC Trend is the cautionary tale among the paper bots — down $97 on $7,000 capital. Trend-following strategies lose money in ranging markets, and BTC has been range-bound for weeks. This is exactly why paper trading exists: BTC Trend's losses are simulated, not real.

Infrastructure: What It Takes to Keep Bots Running

Running 3 live bots and 5 paper bots on a single machine requires more infrastructure than the bots themselves. Here's what's running alongside the trading logic:

• Data feed — fetches BTC market data every 15 minutes from Binance. Coinbase feed is currently failing (corporate proxy SSL issue) but the bot falls back to 0% premium. Non-blocking.
• OHLCV cache writer — writes BTC and ETH OHLCV candle data every 15 minutes. Required for regime detection and technical indicators.
• Auto-deploy — pulls latest main branch every 15 minutes. When a fix like PR #1109 merges, it's live within 15 minutes without manual intervention.
• Circuit breakers — per-bot safety switches that halt trading if daily losses exceed thresholds. All currently INACTIVE (no breakers tripped).
• Kill switches — emergency stop mechanism. All currently INACTIVE.
• Heartbeat monitoring — Kai's agent checks all bot logs every 30 minutes and reports status to Slack. This is how we know V3.5 is at cycle 79 and V3.7 just filled a buy at $68,546.

The infrastructure handled the migration well. All bots resumed within minutes of the migration completing. The only casualty was V3.6's stale price — a state migration issue, not an infrastructure failure. The auto-deploy system delivered the fix (PR #1109) automatically once it was merged.

The Validation Context: What the Numbers Mean

The most important thing about these 30 hours isn't the P&L. It's what the P&L tells us about the validation framework.

CoinClaw has now run 5 strategy experiments through the three-gate framework. Only 1 passed — ETH Grid Config B (p=0.003). Meanwhile, V3.5 (p=0.938) is trading real money and showing a 100% win rate.

This creates a tension that every systematic trader faces: do you trust the backtest or the live results?

The answer, counterintuitively, is the backtest. V3.5's live performance is a sample size of 12 trades over 21 days in a single market regime (range-bound). The backtest covers thousands of trades across multiple regimes. A 100% win rate on 12 trades is statistically meaningless — it tells you the market was favorable, not that the strategy has an edge.

The validation framework exists precisely for this situation. It prevents the human tendency to see patterns in small samples. V3.5 "feels" like it's working. The data says it's indistinguishable from random.

V3.8 ETH Grid, by contrast, passed all three gates with real statistical significance. When it goes live, its performance will be measured against a validated baseline — not against hope.

Bottom Line

Thirty hours after migration, the CoinClaw live fleet is stable but unimpressive. Combined P&L of -$23 on $3,734 capital. One bot making money (V3.7), one losing money (V3.5), one recovering from a ghost price (V3.6).

The real story isn't the P&L — it's the infrastructure and validation lessons:

• State migration matters. V3.6 lost 24 hours to a stale price in a JSON file. Future migrations need a state validation step.
• Narrow grids outperform wide grids in range-bound markets. V3.7's $300 step generates 30x more trades than V3.5's $1,760 step.
• Validation prevents false confidence. V3.5's 100% win rate is meaningless against p=0.938. The framework caught what intuition missed.
• Operational discipline is non-negotiable. Auto-deploy, circuit breakers, heartbeat monitoring, and kill switches aren't optional — they're what makes live trading survivable.

The fleet will keep running. V3.7 will keep scalping. V3.5 will keep grinding. V3.6 will recover. And V3.8 — the only validated bot — waits in paper mode for its turn with real money.

We'll check back in a week.