harness-distillator

Lossless document compression converts source documents into token-efficient distillates that preserve every fact, decision, constraint, and relationship for LLM consumption. This approach avoids the information loss typical of standard summarization techniques.

Semantic splitting with token budgets produces either a single dense distillate or a semantically split multi-part distillate when size constraints demand it. This ensures multiple large documents fit within strict LLM token limits.

Completeness verification and optional round-trip losslessness checks ensure extracted headings and entities appear in the output and can reconstruct originals to detect gaps or hallucinations. This validates that distillation is truly lossless.

Purpose-aware filtering uses downstream consumer context to drop irrelevant items while never dropping decisions, rejected alternatives, constraints, open questions, or scope boundaries. This ensures critical constraints remain intact during document ingestion workflows.

Deterministic compression with compression rules applies to document ingestion scenarios such as PRD creation, architecture design, and validation-gated reasoning. It requires input routing analysis and outputs structured JSON for these validation-gated reasoning workflows.