A Boundary Test for Retrieval Cache Freshness

Validating previously retrieved sources does not establish that current retrieval would return the same evidence: inserted competitors can change top-k without modifying any cached source. We specify a restricted exact-retrieval certificate and audit it on real SciFact and NFCorpus text with controlled snapshot updates. Across 49,840 query/update decisions, the certificate agrees with exact recomputation. With 128 insertions, dependency-only reuse misses changed traces in 21.7% and 30.2% of decisions, respectively. However, classical incremental top-k maintenance is faster than certificate-plus-rescan in that setting. The contribution is a reproducible contract-focused audit, not a new ranking algorithm, production safety guarantee, or answer-quality evaluation. AI assistance. OpenAI-based AI assistants assisted research-question development, literature discovery, argument checking, experiment implementation and execution, analysis, visualization, and manuscript drafting. Exact assistant model versions are not documented in the supplied package. The experiments use real corpus text with controlled simulated snapshot updates; these are not historical production updates. Materials and attribution. The manuscript reports experiments on SciFact and NFCorpus. This proposed deposit contains the manuscript PDF only; it does not redistribute raw corpus text, archives, or the software/data package. Research draft, revision 3. Not peer reviewed. Release status. The uploader approved public release of this revision on 2026-11-08. Author-review and submission-approval notes in the PDFs were written during draft preparation. This is a preprint and has not been peer reviewed.