net/hanzo-hmm

hanzo-hmm — Hidden Markov Model + Hamiltonian MarketMaker

Pure Hidden Markov Model primitives plus a Hamiltonian MarketMaker built on top of them, used in the Hanzo network to price heterogeneous compute.

Naming. This crate is not an LLM engine. It is the pricing / routing layer. The actual model-serving engine lives in ~/work/hanzo/engine (a mistral.rs fork exposed as hanzo-engine).

Two layers, one crate

Layer	Module	Purpose
HMM core	hmm_core	Viterbi, forward-backward, Baum-Welch on HiddenMarkovModel<S, O>
MarketMaker	lib::MarketMaker	Hamiltonian price dynamics + BitDelta adapters + active-inference routing

The HMM layer is standalone and reusable for any sequence modelling task (state detection, sequence prediction, anomaly detection). The MarketMaker layer composes HMM regime detection with Hamiltonian mechanics and BitDelta-quantized per-tenant adapters to set prices and routing decisions across compute classes.

HMM core usage

use hanzo_hmm::HiddenMarkovModel;

let states       = vec!["Fair", "Loaded"];
let observations = vec![1, 2, 3, 4, 5, 6];
let initial      = vec![0.5, 0.5];
let transitions  = vec![
    vec![0.7, 0.3],
    vec![0.4, 0.6],
];
let emissions = vec![
    vec![1.0 / 6.0; 6],
    vec![0.1, 0.1, 0.1, 0.1, 0.1, 0.5],
];

let hmm = HiddenMarkovModel::new(
    states, observations, initial, transitions, emissions,
)?;

let observed = vec![6, 6, 6, 1, 2];
let path     = hmm.viterbi(&observed)?;          // most likely state sequence
let p        = hmm.forward(&observed)?;          // P(observations | model)
let _seq     = hmm.generate(100);                // sample observations

MarketMaker usage

use hanzo_hmm::{MarketMaker, Config, RoutingRequest, UserPreferences, PerformanceRequirements};

let mm = MarketMaker::new(Config::default()).await?;

let decision = mm.route_request("tenant-42", &RoutingRequest {
    input: "...".into(),
    context: vec![],
    preferences: UserPreferences {
        max_latency_ms: Some(1_500),
        max_cost_per_token: Some(0.001),
        preferred_models: vec![],
        quality_threshold: 0.8,
    },
    requirements: PerformanceRequirements {
        min_tokens_per_second: Some(40.0),
        max_memory_gb: Some(48.0),
        requires_function_calling: false,
        requires_vision: false,
    },
    observations: vec![0.1, 0.2, 0.05, 0.4],
}).await?;

Algorithms

• Viterbi — most likely state sequence given observations
• Forward — P(observations | model)
• Backward — backward probabilities per state
• Baum-Welch — parameter learning from observation sequences

let path  = hmm.viterbi(&observations)?;
let prob  = hmm.forward(&observations)?;
let beta  = hmm.backward(&observations)?;
hmm.baum_welch(&training_sequences, max_iterations, tolerance)?;

License

MIT