Question
Can an AI system transmit less text between components while preserving enough evidence for accurate reasoning?
Hypothesis
Context should be task-shaped. Retrieval can send compressed evidence packets when the downstream model needs constraints, entities, citations, or decisions rather than full passages.
Method
Compare raw chunks, extractive summaries, entity tables, and citation-first evidence packets against the same generation task.
Prototype
Build a small pipeline that transforms retrieved documents into typed evidence records with source spans and confidence notes.
Notes
The system should never hide provenance. Compression is useful only when a user can still inspect the original evidence.
Results / Open Questions
Early expectation: entity-heavy tasks benefit most. Open question: how to detect when compression will remove necessary nuance.
References
Placeholder for papers on context compression, retrieval evaluation, and efficient inference.