01 · Declare
Provide source documents, a task, and a token budget—then name the exact evidence spans that must survive.
Context Pack EvaluatorDeclared evidence · deterministic measurement
Measure the context reduction. Test what evidence survived.
Define the passages a workflow must retain, then test a bounded Context Pack against that requirement instead of accepting a compression claim on faith.
The evaluator makes the scope of a result explicit: task, token budget, sources, required evidence, and what did or did not appear in the compiled pack.
Evaluation method
Three things a credible compression result needs.
02 · Compile
The deterministic compiler ranks, deduplicates, and bounds context. No model grades the supplied material.
03 · Report
Receive token and byte metrics, source coverage, duplicate removal, and retained-or-omitted results for every declared span.
Result boundary
Retained is a traceability result, not a truth claim.
Retained means the exact required span appears in an included passage from its declared source. Omitted means it does not. Neither result establishes factual accuracy, model-answer quality, legal compliance, or downstream hallucination behavior.
Use in an assessment
Compare a real workflow before adopting new infrastructure.
Repeat the evaluation on a representative, sanitized corpus. Publish the task, budget, source count, evidence definition, and measured results. Do not generalize one benchmark into a universal savings or quality promise.