PacketCert: Set-Level Evidence Certification for Compact RAG

Compact RAG makes an irreversible decision before generation: which retrieved passages survive the context budget. Passage-wise selection can preserve answer F1 while dropping a required hop, the qualifier governing a claim, or counterevidence already present in the pool. PacketCert certifies the selected set by compiling a query–pool pair into linked proposition, qualifier, and available-counterevidence requirements; learning coverage and selection scores from exact typed-cover targets; decoding witnessed sets under a hard cap; and using a separately parameterized verifier for deficiency-directed expansion or abstention. Independent packet judgments provide the primary validation. On a 600-packet cohort balanced over HotpotQA, MuSiQue, 2WikiMultiHopQA, and ASQA, approval precision/recall is 91.4/88.7%, versus 82.5/77.9% for an identical-safeguard CE+NLI selector; the certificate score reaches AUROC 0.881 and 2.9% harmful-omission risk at 80% coverage, compared with 7.6% for SURE-RAG. On HotpotQA, compiled-requirement coverage is 92.8% at 1.4K evidence tokens, versus 81.7% at 2.1K for CE+NLI. In a prespecified analysis of 36 answer-F1 near ties, approval reverses 14 choices and blinded judges prefer its winner in 11. On all 948 ASQA examples, PacketCert attains 33.3 Disambig-F1 with 5.9% harmful omission at 1.6K tokens. Set-level certification therefore improves compact evidence construction and supplies a human-aligned tie-breaker where answer quality cannot distinguish operating points.