PAKT
Lossless-first prompt compression for structured data — deterministic, model-free, and honest about where it loses.
- Byte-for-byte round-trip at layers L1–L3; lossy L4 is opt-in
- Tokenizer-aware packing measured with a real BPE tokenizer
- Published savings of 27–69% — and a published −25% failure case
- Live-model comprehension eval, p-values included
Structured data is the most compressible thing you send to an LLM — JSON’s ceremony tokenizes terribly. PAKT converts JSON, YAML, CSV, and Markdown into a compact pipe-delimited form that round-trips byte-for-byte. No model in the loop, so there is nothing to hallucinate.
The README publishes the case where PAKT loses — small nested configs get 25% worse — alongside the wins, and its comprehension eval reports the p-value that makes it a weak discriminator. If a compression tool won’t show you its losses, it’s marketing.
Ships as a library, CLI, MCP server, Python wrapper, playground, desktop tray, and browser extension.