# Using Absorbing Markov Chains + EmbeddingGemma 2 as a Decision Engine: ~2ms inference with 100% GTO Parity on Vegas Blackjack — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/02a7a0026c8bc18a

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-10-14 16:55 ingested 2026-10-15 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1x5wjps/using_absorbing_markov_chains_embeddinggemma_2_as/)

## retained text



```
**[UPDATE / RETRACTION: The "100% GTO" claim had a prompt-leakage bug]**
We patched the repo to pass strictly neutral states, here are the real, honest numbers:
• Latency: Still ~2ms (ONNX + Absorbing Markov Chain).
• True Zero-Shot GTO Parity: ~15% - 29% across all 340 canonical hands.


**[Benefits of Tier 0 Absorbing Markov Chain + Embedding model]
on JevBench, Tier 0 achieved 12 / 12 early exits with 100% accuracy at 16.5 ms p50, and on Decision Index 0.3, it safely rejected 344 / 344 out-of-domain items with 0 false exits


Gevva0 Cascading Decision Engine:

**Tier -1 · Static Rules (**`<50 µs`**):** Hot-reloadable regex and keyword guards that run in microseconds without touching the GPU

**Tier 0 · Absorbing Markov Chains (**`<3 ms`**):** Embeds queries into a 768D manifold using EmbeddingGemma-2 (740M) and solves random-walk absorption probabilities in closed form, hitting 100% accuracy on early exits with zero false exits on 344 out-of-domain items.

**Tier 1 · Fast-Path Logits (**`~35 ms`**):** Runs a single forward pass on Gemma 4 26B-A4B (MoE, 4B active) with a custom K=64 micro-head that slices the 262,144-row vocabulary down to just the 52 answer letters cutting output-head compute by 99.98% and resolving 77.9% of queries without generating a single token.

**Tier 2 · Hot-KV Bounded CoT (**`≤ 256 tokens`**):** Escalates the hardest 17.1% of requests into bounded Chain-of-Thought reasoning directly from the KV cache Tier 1 already prefilled (`reuse_kv=True`), eliminating double prompt processing and cutting long-context latency by up to 30%.

**Highlights:**

* JevBench, Gevva0 beats **Claude Opus 4.6 Thinking** on chance-corrected intelligence (**84.62 vs. 83.90**) at a **44 ms median latency** (vs. Opus's 4,850 ms)
* Decision Index 0.3, our 4B-active MoE cascade outperforms dense 27B models (like Kev 27B) at a **60 ms p50**—roughly **7× faster** than 27B alternatives.
* 60 FPS Real-Time Game-Loop Ready  (Pacman demo included)

[https://github.com/solvingSteve/Gevva0](https://github.com/solvingSteve/Gevva0)
```



document fetched 2026-10-15 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5wjps/using_absorbing_markov_chains_embeddinggemma_2_as
