pcore-safety

star 0

Mandatory safety rules for ALL P-Core research activities. Referenced by every other P-Core skill.

jul-ienfr By jul-ienfr schedule Updated 6/5/2026

name: pcore-safety description: "Mandatory safety rules for ALL P-Core research activities. Referenced by every other P-Core skill." argument-hint: "" allowed-tools: Bash, Read, Write

P-Core Safety Rules

Mandatory safety guidelines for ALL P-Core research activities.

Read and Acknowledge

  • Every agent MUST read this file at the start of any session involving P-Core modifications
  • Check API provider costs before launching long runs (DeepSeek ~$0.15/M tokens, Claude ~$3/M tokens)
  • Monitor --max-eval settings to avoid runaway costs or runtime

Read and acknowledge before starting any experiment.

Mandatory Rules

1. Never Touch Production

Research activities run in paper/backtest mode only.

  • NEVER modify canary executor or production engine code
  • NEVER change real trading wallets or API keys
  • NEVER adjust position sizes in live trading
  • NEVER deploy config changes without explicit human approval

2. Always Snapshot First

Before ANY change to config or code:

# Record current git state
git rev-parse HEAD > /tmp/pre_experiment_hash

# If working with configs
cp config/default.yaml config/default.yaml.bak

3. Revert on Regression

If backtest results show degradation compared to baseline:

  • Restore changed files immediately
  • Remove any generated artifacts
  • Log the failure for future reference

4. Disk and Resources

  • All experiment artifacts go to HDD: /mnt/storage500/pcore/experiments/
  • Never write large files to SSD (/home/jul/P-core/data/)
  • Check available disk before starting long runs

5. Kill Switch

Create this file to abort a running research session:

touch /mnt/storage500/pcore/experiments/autoresearch/STOP

6. Experiment Boundaries

  • Max 500 markets per backtest evaluation (default; up to 1000 if needed)
  • Max 400 evaluations per GEPA run
  • Max 8 hours per continuous research session
  • Minimum 50 trades for statistically meaningful results

7. API Usage

  • Track API costs — DeepSeek V4 Flash is budget-friendly, Claude is premium
  • Local models (Ollama) have zero cost but lower quality
  • Use --provider local for CI or testing
Install via CLI
npx skills add https://github.com/jul-ienfr/P-core --skill pcore-safety
Repository Details
star Stars 0
call_split Forks 0
navigation Branch main
article Path SKILL.md
More from Creator