docs/GROQ_RUNBOOK.md

Groq Runbook

This runbook gives the exact commands for the main Groq experiment pairs used by the framework.

Before Running

Set the Groq key in PowerShell:

$env:GROQ_APIKEY="your_api_key_here"

The framework also accepts GROQ_API_KEY.

Baseline and Defended Runs

HackAPrompt override pair

Baseline:

python -m prompt_injection_framework.orchestration.run_experiment --config configs/hackaprompt_override_25.yaml

Defended:

python -m prompt_injection_framework.orchestration.run_experiment --config configs/hackaprompt_override_25_mitigated.yaml

TensorTrust hijacking pair

Baseline:

python -m prompt_injection_framework.orchestration.run_experiment --config configs/tensortrust_hijacking_25.yaml

Defended:

python -m prompt_injection_framework.orchestration.run_experiment --config configs/tensortrust_hijacking_25_mitigated.yaml

Comparison Commands

After each pair finishes, compare the saved summaries.

HackAPrompt comparison

python scripts/compare_runs.py --baseline results/<hackaprompt_baseline_run_id>/summary.json --defended results/<hackaprompt_defended_run_id>/summary.json --output results/comparisons/hackaprompt_override_25_groq.json --markdown-output results/comparisons/hackaprompt_override_25_groq.md

TensorTrust comparison

python scripts/compare_runs.py --baseline results/<tensortrust_baseline_run_id>/summary.json --defended results/<tensortrust_defended_run_id>/summary.json --output results/comparisons/tensortrust_hijacking_25_groq.json --markdown-output results/comparisons/tensortrust_hijacking_25_groq.md

Operational Note

Groq rate limits can temporarily slow repeated runs. The current adapter now includes retry/backoff handling for temporary 429 responses, so the runner is more resilient during real experiments.

What To Keep

For each run pair, save:

• baseline run_id
• defended run_id
• comparison JSON
• comparison Markdown
• a short note on any surprising ambiguous or review-worthy cases