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.

1. Pick a Test Scenario 5 Cases 2. Triple-Guard Tripwires 2.1 ms Guard 1: Negation & Syntax PASS 0.0 ms check for "do not", "instead of", binary yes/no Guard 2: Manifold Proximity OOD PASS Requires d min ≤ 0.22 to nearest landmark Guard 3: Entropy & Margin Gate PASS Requires Entropy H ≤ 0.45 & Margin Δ ≥ 0.25 ❌ Naive Cosine Similarity Direct Line Fooled by keywords ✅ Tier 0 Hybrid AMC Maze Absorption 65% Cos + 35% AMC EXIT 0: LOCKED VERDICT 1.9 ms