Results Format

Each SemaTune run writes into results_dir. Treat the directory as an audit trail for the control loop: the target output, per-window metrics, tuner response, accepted parameter values, and LLM prompts or responses when used.

Important files:

• optimization_history_<benchmark>_<timestamp>.json: full optimization history
• window_<n>_metrics.json: system metrics per window when collected
• target-specific window output directories

For LLM runs:

• llm_api_logs/llm_responses.jsonl: one compact record per LLM response, including model, agent, parsed response, response text, and a pointer to the full log file
• llm_api_logs/llm_api_iter<iteration>_<agent>_<timestamp>.txt: exact prompt, generation config, response text, parsed response, and raw SDK representation

For sysbench_cpu runs:

• sysbench_cpu_windows/continuous_sysbench.log: the live sysbench --report-interval stream for the whole run
• sysbench_cpu_windows/window_<n>/sysbench_intervals.json: parsed interval reports and the per-window aggregate used by the optimizer
• sysbench_cpu_windows/window_<n>/sysbench_interval.log: raw interval lines captured for that window

The optimizer no longer writes run_metadata.json; it also does not capture git diffs or dirty status. Keep experiment notes outside the result directory when you need lab-book metadata.

Use Reading Results for example JSON snippets and jq commands.