/experiment

Run a standard experiment iteration. This command analyzes your current code, executes the training run, logs results, and updates the experiment tracker.

Usage

/experiment

There is also a bolder variant:

/experiment-bold

The -bold variant uses more aggressive hyperparameter changes and architectural modifications.

What It Does

1. Reads project context -- Scans your codebase to understand the current model architecture, dataset, and training configuration.
2. Plans the iteration -- Determines what to change based on prior experiment results and insights.
3. Executes the run -- Applies changes and runs the training script.
4. Logs results -- Appends a new row to experiment_log.tsv with the configuration, metrics, and outcome.
5. Updates insights -- If the run produces a notable finding, appends to insights.md.

Example

> /experiment

After running, the experiment log will contain a new entry:

2025-04-15T10:32:00Z    exp-042 success Increased learning rate to 3e-4 {"loss": 0.342, "accuracy": 0.891}  {"lr": 3e-4, "batch_size": 32}

Artifact Format

The command writes to:

• experiment_log.tsv -- One new row per run (tab-separated)
• insights.md -- Appended only when the run yields a meaningful learning

/experiment vs /experiment-bold

Use /experiment for incremental, safe iterations. Use /experiment-bold when you want to explore further from the current best configuration -- larger learning rate jumps, different architectures, or unconventional approaches.