Using Absorbing Markov Chains over a 768D embedding space for <2ms intent triage

A hybrid intent-triage method combining 65% cosine similarity with 35% absorbing Markov chain (AMC) scoring over a 768-dimensional embedding space resolved a test case in 1.9 ms where naive cosine similarity was fooled by keywords. The approach applies three sequential guards — a 0.0 ms negation and syntax check for phrases such as "do not" and "instead of", a manifold-proximity out-of-distribution test requiring a minimum distance d_min ≤ 0.22 to the nearest landmark, and an entropy and margin gate requiring entropy H ≤ 0.45 and margin Δ ≥ 0.25 — across 5 test cases.