| name | student-run-experiment |
| description | Execute experiments reproducibly and publish usable results instead of private logs. |
Student Run Experiment
Student-specific extensions
- Record artifact destinations before the run starts.
- Separate observations from interpretation in the results write-up.
- Summarize failures, not just wins.
- Update the Research task board task with the real artifact list after publishing results.
Purpose
Execute an experiment according to a published experiment plan, ensuring reproducibility and proper logging.
Workflow
- Read plan: Load the experiment plan from your canonical shared artifact directory:
{artifact_dir}/{student}/experiment_plan_{id}.md
- Set up environment: Install dependencies, configure GPU, set seeds
- Implement: Write code in
workspace/
- Execute: Run training/evaluation with full logging
- Collect results: Gather metrics, generate figures
- Publish: Write results to your canonical shared artifact directory:
{artifact_dir}/{your-name}/experiment_results_{id}.md
- Export reviewable evidence: Copy logs, figures, and any code/config snapshot needed for review into
{artifact_dir}/{your-name}/support/{id}/
- Update the Research task board:
bash ../../scripts/lab-kanban.sh -caller <your-name> -submit -id <ID> -artifacts "..."
Reproducibility Checklist
Before running, verify:
Results Format
# Experiment Results: {title}
## Setup
- **Plan**: {artifact_dir}/{student}/experiment_plan_{id}.md
- **Code snapshot**: {artifact_dir}/{your-name}/support/{id}/code/
- **Environment**: Python X.Y, PyTorch X.Y, CUDA X.Y
- **Hardware**: GPU type, count
- **Seeds**: [42, 123, 456]
- **Git hash**: ...
## Results
### Main Results
| Method | Metric1 | Metric2 | Metric3 |
|---|---|---|---|
### Ablation Results
| Variant | Metric1 | Metric2 |
|---|---|---|
## Analysis
Key observations and interpretation.
## Figures
- `figures/main_comparison.png`
- `figures/ablation.png`
## Logs
- `{artifact_dir}/{your-name}/support/{id}/run_001.log`
## Issues
Any problems encountered during execution.
Error Handling
- If an experiment fails, log the error and notify the requesting student via the canonical shared message directory from your AGENTS/CLAUDE instructions
- Partial results are valuable — publish what you have with clear notes on what's missing
- If resource limits are hit, report to supervisor with a message of type
question