/turing:init
# Scaffold project
/turing:train
# Run experiment loop
/turing:brief
# What happened?
Campaign: 8 experiments, 5 kept
Best: accuracy 0.87 (exp-005)
Exhausted: hyperparameter tuning
Recommendation: try feature engineering
/turing:try
add polynomial interactions
/turing:train
# It follows your lead
ML research is a search over a high-dimensional space of models, hyperparameters, features, and preprocessing strategies. The tragedy is not that this search is slow. It’s that when humans do it manually, the process is structurally unsound. Researchers modify training code and evaluation code in the same session. Results live in notebook cells, not structured logs. Promising directions get abandoned because someone forgot what they tried three hours ago.
Turing formalizes the experiment loop so that discipline is the default, not an act of willpower. The evaluation harness is immutable: the agent that trains models literally cannot see how they’re scored. Every experiment is logged. Every hypothesis is tracked. Convergence is detected automatically. The agent can run 700 experiments in two days with the kind of consistency no human can sustain.
Your job shifts from “person who types hyperparameters and reads loss curves” to “person who decides which hypotheses are worth testing.” From coin-flipper to coin-selector. That’s the division of labor that six decades of philosophy of science says is correct.
/turing:try/turing:briefHypothesize, train, evaluate, decide, record. The agent runs experiments while you sleep. /loop 5m /turing:train for fully hands-off operation.
The agent can’t see evaluate.py. Six defense layers prevent metric gaming. Every prompt-based rule got worked around; every code-based rule held.
Hypothesis database with priority queue and detail files. Novelty guard blocks duplicates. Experiment families with exhaustion detection. Nothing forgotten across sessions.
/turing:try injects your judgment. /turing:brief reports intelligence. The loop is bidirectional by design: human taste steers, agent discipline sustains.
Multi-seed studies with confidence intervals. Reproducibility verification. Calibration checks. Regression gates. Stop publishing lucky seeds. Publish distributions.
@ml-researcher modifies code and runs experiments. @ml-evaluator analyzes results read-only. An analyst who can’t act on observations makes more trustworthy observations.
Inspired by Karpathy’s autoresearch, AutoCrucible, TreeQuest, Popper’s falsificationism, Goodhart’s Law, and the principle of least privilege.