The Retrieval Gate Reads the Page Before the Generator Does

A released four-billion-parameter decision model answers typed questions about retrieved HTML without generating tokens, making it a screening layer for prompt injection and synthetic-page pollution. On 1,000 forum-style questions, attack success fell from 27.9 percent with no defense to 13.6 percent with published mitigations and 5.7 percent with the gate; four of 186 benign instruction-bearing pages were wrongly flagged. In a separate test where two thirds of candidate pages were fake, filtering before top-ten selection held exposure to 0.5 percent and accuracy to 83 percent, compared with 28 percent accuracy without filtering. These results are benchmark-specific, but they favor gating the pool before retrieval commits.