The proliferation of publicly accessible large language models (LLMs) intensifies the need for trustworthy provenance tracking and covert communication channels. We present SemBits, a lightweight, model-agnostic watermarking framework that embeds arbitrary binary payloads into free-form text while preserving fluency and stylistic diversity. SemBits constrains autoregressive decoding with a sentence-level accept–reject loop guided by semantic locality-sensitive hashing (LSH). Each newly generated sentence is hashed in the embedding space, and it is accepted only when its bucket index falls inside a green list derived from the generation prefix and the secret message. This strategy requires no modification or fine-tuning of the underlying model and adds merely \(7\,\%\) sampling latency with a perplexity overhead below \(0.2\) . Experiments show that SemBits achieves \(99.3\,\%\) message-recovery accuracy for \(7\) -bit payloads hidden in \(200\) -token passages produced by OPT 1.3B and Llama-3-8B. The watermark survives up to \(10\,\%\) token-level paraphrasing, moderate synonym substitution, and sentence reordering, outperforming lexical watermark baselines by a factor of \(16\times \) in false-positive reduction and nearly doubling robustness to semantic perturbations.