Absorbing Markov Chains Achieve <2ms Latency on Cached Candidates for Zero-Token Intent Triage

Methodology

Direct cosine similarity often misroutes queries due to surface keywords. The proposed Semantic Triage engine projects queries onto a 60-node domain manifold and solves an Absorbing Markov Chain ($B = (I - Q)^{-1} * R$) on the GPU. This approach traces semantic paths (e.g., Idempotency Key -> Replica Write -> WAL Lock) to absorb into the correct domain without invoking an LLM.

Performance Benchmarks

The JevBench results highlight a significant efficiency gain for the zero-token approach compared to LLM-based tiers:

• Tier 0 (AMC + EmbeddingGemma2): 22ms p50 latency, 100% accuracy (11/11 items).
• Tier 1 (Gemma4 26B A4B): 42ms p50 latency, 91.49% accuracy (172/188 items).
• Tier 2 (Gemma4 + CoT): 1,560ms p50 latency, 75.61% accuracy (31/41 items).
• 4-Tier Cascade (Gevva0): 46ms p50 latency, 89.17% accuracy (214/240 items).

Latency and Safety

Latency varies by context: ~1.5ms for short prompts with cached candidates versus ~22ms for embedding unseen dynamic options on 400-token contexts. The system includes three safety guards: Negation & Syntax, Manifold Proximity (OOD), and Entropy & Margin Gate.