# Work we publish — Hanzo AI > Hanzo builds infrastructure that has to hold up under audit, regulation and adversaries. Research # Work we publish Hanzo builds infrastructure that has to hold up under audit, regulation and adversaries. [Explore papers.hanzo.ai ↗](https://papers.hanzo.ai)[Specifications (HIPs) ↗](https://hips.hanzo.ai)[Benchmarks](https://hanzo.ai/benchmarks) papers.hanzo.ai - Kai: Calibrated Discrete Decision Intelligence for Autonomous Agent Loops01 - Linkage, Not Similarity02 - Hanzo AI Chain (AIVM)03 - Cloud Economics of a Sovereign OSS Stack04 - One Native Stack for Private, Continuously-Learning AI05 - Native ROCm Inference on a Consumer RDNA3.5 APU06 Oct 2026. Factorized Discrete Choice, Calibrated Distributions, and Deterministic Policy Governance Sep 2026. What long-term agent memory has to retrieve, and how it is measured Aug 2026. Useful-Work Mining, the Native AI Coin, and Post-Quantum Omnichain Settlement Jun 2026. A Regulated Capital-Markets Migration Case Study Jun 2026. Inference at the Bandwidth Wall, On-Device QLoRA, One Engine Jun 2026. Reaching llama.cpp Decode Parity via a Unified 1-bit-to-Full Quant Core ## Papers Each one links to its page on papers.hanzo.ai, where the PDF and the LaTeX live. [Kai: Calibrated Discrete Decision Intelligence for Autonomous Agent LoopsOct 2026Factorized Discrete Choice, Calibrated Distributions, and Deterministic Policy Governance](https://papers.hanzo.ai/kai-decision-model/) [Linkage, Not SimilaritySep 2026What long-term agent memory has to retrieve, and how it is measured](https://papers.hanzo.ai/hanzo-context-memory/) [Hanzo AI Chain (AIVM)Aug 2026Useful-Work Mining, the Native AI Coin, and Post-Quantum Omnichain Settlement](https://papers.hanzo.ai/hanzo-ai-chain/) [Cloud Economics of a Sovereign OSS StackJun 2026A Regulated Capital-Markets Migration Case Study](https://papers.hanzo.ai/hanzo-cloud-economics/) [One Native Stack for Private, Continuously-Learning AIJun 2026Inference at the Bandwidth Wall, On-Device QLoRA, One Engine](https://papers.hanzo.ai/hanzo-native-stack-thesis/) [Native ROCm Inference on a Consumer RDNA3.5 APUJun 2026Reaching llama.cpp Decode Parity via a Unified 1-bit-to-Full Quant Core](https://papers.hanzo.ai/hanzo-rocm-inference/) [Edge LLM Inference Across Commodity AcceleratorsJun 2026Measured AMD, NVIDIA, and Apple Performance](https://papers.hanzo.ai/hanzo-cross-backend-inference/) [Native Training in the Hanzo EngineJun 2026One Engine, One Quant Format, One Backend](https://papers.hanzo.ai/hanzo-native-training/) [Continuously-Learning Private AIJun 2026Self-Improvement and Privacy in the Hanzo Native Stack](https://papers.hanzo.ai/hanzo-continuous-learning-privacy/) [Economical Serving and the hanzo.ai PlatformJun 2026A Bandwidth-Optimal Unified Train-and-Serve Runtime](https://papers.hanzo.ai/hanzo-platform-infra-desktop/) [The Unified Sovereign-Tenant CloudJul 2026Linear Shared-Nothing Scaling via Per-Tenant SQLite and In-Process Composition](https://papers.hanzo.ai/hanzo-unified-tenant-cloud/) [Refutation-Driven Performance EngineeringJul 2026An Empirical Study of a Multi-Week GPU Kernel Campaign](https://papers.hanzo.ai/refutation-driven-perf/) [Evolutionary Schedule Search in a One-Source Kernel DSLJul 2026And the Continual-Improvement Loop It Anchors](https://papers.hanzo.ai/evolutionary-schedule-search/) [Hanzo RouterJul 2026Memory-Aware, Local-First LLM Routing with a Learned SLO-Constrained Policy](https://papers.hanzo.ai/hanzo-router/) [What a Dormant Agent CostsSep 2026State, Execution and Isolation Measured Separately at a Fleet of One Million](https://papers.hanzo.ai/hanzo-dormant-agents/) [Hops Do Not Fix Multi-Hop RetrievalSep 2026A Negative Result on Long-Horizon Conversational Memory, and What the Failure Actually Is](https://papers.hanzo.ai/hanzo-multi-hop-retrieval/) [Retrieval Scoped by SubjectSep 2026A factorial measurement of conversational memory on LoCoMo, and the negative controls that locate the effect](https://papers.hanzo.ai/locomo-subject-scope/) [Hanzo Network WhitepaperOct 2024L1 Blockchain for Decentralized AI Compute](https://papers.hanzo.ai/network-whitepaper/) ## Specifications HIPs. The index is hips.hanzo.ai. These are the ones the families are written against. [All HIPsThe specification index.](https://hips.hanzo.ai) [HIP-0000The architecture framework.](https://hips.hanzo.ai/docs/hip-0000-hanzo-ai-architecture-framework) [HIP-0510Enso, the learned router.](https://hips.hanzo.ai/docs/hip-0510-enso-learned-router) [HIP-1332Kai, the decision model.](https://hips.hanzo.ai/docs/hip-1332-kai-the-decision-model) [HIP-0039Zen model architecture.](https://hips.hanzo.ai/docs/hip-0039-zen-model-architecture) [HIP-1334The research runtime.](https://hips.hanzo.ai/docs/hip-1334-research-unified-runtime) [HIP-1145How an experiment is recorded.](https://hips.hanzo.ai/docs/hip-1145-research-experiment-record) ## Standards The standards the runtime, the gateway, and the benchmark arena are built to. [HIP-0043The inference engine.](https://hips.hanzo.ai/docs/hip-0043-llm-inference-engine-standard) [HIP-1211The model API.](https://hips.hanzo.ai/docs/hip-1211-ai-the-model-api) [HIP-0004The gateway.](https://hips.hanzo.ai/docs/hip-0004-llm-gateway-unified-ai-provider-interface) [HIP-0019Tensor operations.](https://hips.hanzo.ai/docs/hip-0019-tensor-operations-standard) [HIP-1105The benchmark arena.](https://hips.hanzo.ai/docs/hip-1105-benchmark-measurement-arena) ## Open source The systems described in these papers are open source. The cloud, the gateway, the agent runtime and the infrastructure around them are developed in the open at [github.com/hanzoai](https://github.com/hanzoai), and the platform pays the open-source authors whose work runs on it.