We demonstrate reliable, fine-grained control over language model generation by directly injecting, suppressing, and composing human-interpretable concepts at inference time.