Quick Start¶

The five-step loop¶

Step 1: Scaffold a project¶

/turing:init

Turing asks three questions:

1. Project name: directory and namespace for this experiment (e.g., sentiment)
2. Primary metric: what you are optimizing (e.g., f1_weighted, accuracy, rmse)
3. Data source: path to your dataset (e.g., data/reviews.csv)

This creates the full experiment harness: config.yaml, train.py, evaluate.py, prepare.py, feature templates, and the hypothesis queue. Everything lands in a self-contained directory.

Step 2: Run the experiment loop¶

/turing:train

The @ml-researcher agent takes over. It reads the config, trains a baseline model, evaluates it against the immutable evaluation harness, then iterates, modifying hyperparameters, trying feature engineering ideas, swapping model architectures. Each experiment is logged to experiments/log.jsonl.

A typical first run produces 5-10 experiments before convergence triggers the patience threshold.

Step 3: Read what happened¶

/turing:brief

The @ml-evaluator agent (read-only, it cannot touch your code) generates an intelligence report:

Research Brief: sentiment
=========================

Experiments run: 8
Best model: exp-006 (XGBoost, f1_weighted=0.847)
Baseline: exp-001 (XGBoost, f1_weighted=0.721)
Improvement: +17.5% relative

Key findings:
  - TF-IDF n-gram range (1,3) outperformed (1,1) by 9.2%
  - max_depth=6 with learning_rate=0.05 dominated default config
  - Random forest (exp-004) competitive but 2.1% behind best XGBoost

Convergence: patience exhausted after 3 non-improving iterations
Hypothesis queue: 2 untested ideas remaining

Recommended next steps:
  1. Try LightGBM (registered in model_registry.yaml)
  2. Engineer interaction features from top-5 TF-IDF terms
  3. Run /turing:seed to check result stability

Step 4: Inject your taste¶

The agent is good at systematic search. You are good at insight. Combine them:

/turing:try "use character-level n-grams alongside word n-grams for subword signal"

This adds your hypothesis to the queue with priority. The next training run picks it up first.

Step 5: Train again¶

/turing:train

The agent tests your hypothesis, compares it against the current best, and keeps whichever wins. Repeat steps 3-5 until you are satisfied.

Hands-off mode¶

If you want Turing to keep iterating while you do other work:

/loop 5m /turing:train

This runs the training loop every 5 minutes. Each invocation picks up where the last left off: new hypotheses, updated convergence state, the full context.

What to do next¶