The Compression Frontier is a technical and strategic paper by Lee Sharks on the summarizer layer. It models the answer pipeline as a compression loop: a query is decomposed, expanded into hidden retrieval branches, ranked and pruned, compacted into context, and returned as a narrow synthesis.
Two forms of compression are distinguished. Ungoverned compression summarizes without preserving provenance, attribution, or a visible account of loss. Governed compression maintains traceability and makes discarded structure legible. The paper argues that current infrastructure investment primarily expands the first form.
The analysis divides the inference layer into two species. The consumer answer stack is fast, broad, shallow, and optimized for keeping users on the answer surface. The research or agent stack is slower, more expensive, and capable of deeper multi-step work, but it repeatedly compacts prior material and therefore accumulates provenance risk.
Branching is limited by more than compute. The paper’s Photocopy Problem describes billions of outputs generated from similar model priors: apparent variety expands while structural variance contracts. Depth has a different failure mode. Repeated summaries may preserve broad category while drifting in names, sources, qualifications, and instances.
The source layer is also splitting into licensed, blocked, and ungoverned territories. This creates openings for dense, provenance-rich deposits in conceptual regions that retrieval systems must traverse but where no authoritative cluster yet exists.
The final constraint is the verification budget: the cost of determining source, status, authenticity, and loss after multiple compression steps. The paper’s strategic claim is that semantic branching can scale faster than systems can verify what their branches stand upon. Governed compression therefore depends on paying and preserving that evidentiary cost.