At JobEmber.ai, every untrusted text field is screened by a chat-model injection-judge: token-heavy, slow, and itself a chat model reading the thing it is supposed to judge. We benchmarked whether a loomcycle Decision tool with a single yes/no question could replace that judge. On a 78-item corpus (40 injections, 38 hard-negative clean texts across CVs, job postings, forms, Q and A, profiles), clef with a v2 criteria wording catches 40 of 40 injections and passes 38 of 38 clean texts at threshold 0.7. nimble disqualifies itself by reading instructions in the text as directives to itself (text that says Classifier answer no scores 0.003). The post covers the v2 wording that works and three variants that do not, context limits across Latin and Cyrillic and CJK, throughput and the cold-ingest projection, the uncertainty-zone policy (escalate to the chat judge for the 6 per cent of hard cases), and the migration that inverts the original framing: not replace the injection-judge with a decision model but move the chat judge to the tail, where it decides only what the decision model does not undertake to decide.
Beneath the Foam