Skip to main content
This is the launch playbook for a low-budget deployment. You do not need a high-end Nvidia server to run NOVOSKY live. Keep retraining centralized, keep runtime machines simple, and launch in small steps.
Operating model: the live account and client machines should only pull approved model revisions from R2 and run trading. Retraining, optimization, and model publishing stay in the developer/operator lane.

​
Before Monday

  1. Freeze the approved model revision and config profile set.
  2. Confirm the active profile for each account or client.
  3. Verify the hard drawdown halt, circuit breaker, and Telegram alerts.
  4. Run a dry-run on the exact machine that will trade live.
  5. Confirm PM2 restart behavior and halt behavior.
  6. Make sure you know the rollback revision.

​
Go / no-go gate

Do not launch until all of these pass:
  • Model compatibility check passes.
  • Dry-run completes multiple cycles without crashing.
  • OOS backtest is within the chosen profile’s risk limits.
  • MT5 API connection is stable.
  • Telegram alerts arrive correctly.
  • Max drawdown halt is active and correct.
  • You know how to revert to the last approved model revision.
If one critical check fails, stop and fix it before going live.

​
Monday launch sequence

1

Pull the approved revision

python ml/r2_hub.py --pull --revision vYYYYMMDD-p3
Use the revision that matches the approved profile for the account.
2

Verify compatibility

python3 -c "
import json
mc = json.load(open('models/model_compat.json'))
ml = json.load(open('ml_config.json'))
assert mc['feature_count'] == len(ml['features'])
print('OK:', mc['feature_count'], 'features')
"
3

Run a dry-run

python trading.py --dry
Let it cycle long enough to confirm signal flow, filters, and logging.
4

Start live trading

python trading.py
Monitor the first session closely. Do not leave it unattended on the first day.

​
Low-budget setup

Use the cheapest machine that can reliably run the live bot:
  • 4 vCPU / 8 GB RAM is usually enough for the trading runtime.
  • CPU-only machines are fine for live execution.
  • Keep retraining on a separate developer machine or cloud job.
  • Clients should never need to train models locally.

​
Rollout plan

  1. Launch your own account first.
  2. Watch it for 24-48 hours.
  3. Add one or two trusted client accounts.
  4. Expand only after the canary run is stable.

​
Rollback plan

If the launch misbehaves:
  1. Stop live trading.
  2. Pull the previous approved revision.
  3. Re-run dry-run.
  4. Restart only after the issue is understood.
The goal is predictable execution, not maximum aggressiveness on day one.