mirror of
https://github.com/hanzoai/papers.git
synced 2026-10-22 17:20:17 +00:00
zeekayandHanzo Dev
bc1452e9f0
zenlm/papers had a sweeping withdrawal pass, 5deeeaa, and the owner reverted it in f34f609 for asserting more than it verified. That objection sets the rule here: a claim changes only where the proof is in the paper itself or in arithmetic, never because a run could not be found. Twenty-seven of the eighty-one zen papers meet that bar. Every correction quotes a number the paper already carries or one derived from its own numbers by a division a reader can repeat. Twelve papers led with a figure that appears in no row of the table it summarises. zen-agent averaged GAIA 83.1/68.4/42.3 to 71.2 while its four baseline rows are each the plain mean of their own three levels; the mean is 64.6. zen-embeddings-retrieval and zen3-embedding claim MTEB averages of 72.4 and 74.3 over tables that supply their task counts, where the 56-task weighted means are 67.1 and 67.4. zen-mathematical-reasoning opened with five figures, none of them any row, and explained one as best-of-32 sampling when its own sampling table puts best-of-32 at a different number. zen-dso-protocol claimed 94% of centralized quality and a 78% bandwidth cut against tables reading 99.4% and 99.8%-and-better, the second of them the table that claim captions. zen-synthetic-data claimed an 8.4-point MT-Bench gain on a scale that runs to 10, from a 7.84 baseline, where its table says 0.88. zen-3d quoted the Chair column as the ShapeNet result. zen-multilingual understated its own XNLI average. Also zen-hallucination-reduction, zen-distributed-training, zen-finetuning, zen-legal-ai and zen-coder-flash. Eight more are settled by arithmetic against the hardware or the parameter count. zen-nano, zen-eco and zen-scribe each report INT4 needing as much memory as FP16 or more, while their own FP16 and INT8 rows are exactly two and one bytes per parameter; zen_family_overview already carried the right values. The two zen-designer papers size a stated 235B model at 220 GB in FP16, which is a 110B model, and zen-quantization puts the same model at 470 GB. zen-omni and zen-next recommend cards smaller than the memory the same row requires. zen-live captions a latency table to an A10G when its own deployment table assigns that 98 ms to the H100. zen-hardware-optimization calls a throughput table FP8+INT4 when its cumulative table reaches the identical 31,200 tok/s through INT4 + FP16. Six papers share one template whose visual benchmark block bundles two metrics computed over images a model generates with one computed over text a model emits about an image. No model produces all three, and the block went to four papers across four modalities, so each received at least one metric it cannot compute: VQA v2 and a token rate for two diffusion models, FID and CLIP score for two vision-language models. Each now keeps the rows its own stated output type can produce. Throughput is the one category where the defect is the missing condition rather than the number. Decode reads the active weight set once per token, so a single stream cannot outrun bandwidth over footprint, and zen4-mini, zen3-nano and zen-coder-flash all report above their own bound. Batched serving would make those figures legal and none of the three records a batch size, so the numbers stand and each table now carries its bound and the arithmetic behind it. AIME gets the same treatment in zen-mathematical-reasoning: a score that is not a multiple of 1/30 on a 30-problem exam is legal under averaging, and the caption now says which row has its sample count on record and which do not. What was examined and left is in LLM.md with the reason in each case. Training fleets absent from this estate prove nothing about models trained on rented capacity. zen-reranker's 98% semantic preservation is uncited rather than contradicted. zen-safety-evaluation and zen-vision-architecture state no weighting, so their averages are not determined. zen4-coder-flash puts FP8 on an A100, and Ampere has no FP8 tensor cores, but its claims there are footprint claims and FP8 weight-only storage does run on Ampere. All twenty-seven rebuild under latexmk -Werror. Nineteen have diverged from their counterparts in zenlm/papers and six exist only here, so those fixes do not reach the other site. zen-3d and zen-agent were byte-identical across the two repositories until now, and are the two to mirror first. Co-authored-by: Hanzo Dev <[email protected]>
Hanzo AI Research Papers
Technical papers covering Hanzo AI compute infrastructure, agent frameworks, ML inference, platform services, and defense applications. Includes the Zen model family co-developed with Zoo Labs Foundation.
About
Hanzo AI is an American-owned, Techstars-backed ('17) applied AI lab, founded 2014, building the Zen frontier model family alongside a vertically integrated, post-quantum AI and cryptography stack. The work spans frontier models, an agentic runtime, edge inference, and production post-quantum cryptography — engineered to run sovereign and disconnected in defense, regulated, and air-gapped environments rather than on centralized cloud. The platform these papers document is generally available at hanzo.ai.
2026 Highlights
Each claim below is developed in the catalogued papers (INDEX.md):
- Defense & Reindustrialization suite (16 papers + one-pager). Positioned for the
American industrial base — sovereign AI, robotics, post-quantum, and low-latency
on-chain coordination for factories, vehicles, drones, ships, and contested-edge
systems. Strategy & case studies: sovereign-vs-closed AI (
hanzo-vs-claude), the ~90% cloud-cost migration on real numbers (hanzo-cloud-economics), agentic engineering at regulated scale (hanzo-agentic-prowess), FHE-native confidential computing (hanzo-fhe-native), automated security auditing (hanzo-security), and audit-ready compliance (hanzo-compliance). Capability depth: autonomous agentic AI (hanzo-agentic-ai), DDIL/tactical assured networking (hanzo-assured-networking), FHE-mediated cross-domain MLS (hanzo-cross-domain), edge AI/ML (hanzo-edge-ai-ml), counter-exploitation (hanzo-countermeasures), full-spectrum cyber (hanzo-full-spectrum-cyber), quantum sensing/QKD (hanzo-quantum-sensing-qkd), verification & validation (hanzo-test-evaluation), and the ZAP protocol (hanzo-zap-protocol). Start with the one-pager (defense/hanzo-overview.pdf); the full suite is published at papers.hanzo.ai. - Engine / AI infrastructure (6 papers, measured on-device). A native Rust
train-and-serve engine benchmarked on real silicon: ROCm decode parity with llama.cpp
at 97% of the memory-bandwidth wall (
hanzo-rocm-inference), measured AMD/NVIDIA/Apple edge inference (hanzo-cross-backend-inference), native training inside the inference stack (hanzo-native-training), continuously-learning private AI (hanzo-continuous-learning-privacy), serving economics (hanzo-platform-infra-desktop), and the unifying thesis (hanzo-native-stack-thesis). Unusually honest — they name their own stubs and reverted kernels and claim no win over NVIDIA. - Unified sovereign-tenant cloud (measured). The whole multi-tenant cloud as one Go
binary — 56 in-process subsystems over shared-nothing storage, each tenant its own
post-quantum-encrypted SQLite file. Measured on the production open path: ≈0.25 MB RAM
per actively-served tenant, near-zero idle — memory scales with concurrency, not
registrations, so one billion registered tenants at 1% peak concurrency fits ≈20
commodity 128 GB nodes (
hanzo-unified-tenant-cloud; architecture companioncloud-unified-binary). - Production post-quantum cryptography. All three NIST PQC standards — FIPS 203
(ML-KEM), 204 (ML-DSA), 205 (SLH-DSA) — integrated into a live consensus system with
hybrid classical/post-quantum security (
hanzo-full-spectrum-cyber). - Formal verification. 50 machine-checked Lean 4 proofs, 18 cryptographic security
reductions (incl. ML-DSA EU-CMA, ML-KEM IND-CCA2), and NIST Known-Answer-Test
validation across FIPS 203/204/205 (
hanzo-formal-verification). - ZAP (Zero-copy Application Protocol). 2.2 ns parse latency, 11.5M transactions/s,
91% memory and 96% energy reduction vs JSON-RPC; native post-quantum transport
(ML-KEM-768 + ML-DSA-65) and W3C DID identity (
hanzo-zap-protocol). - Agentic runtime. 83 specialized agents and 15 orchestrators executing 200–300
sequential tool calls, fully offline on embedded ARM with quantized Zen models
(
hanzo-agentic-ai,hanzo-edge-ai-ml). - Zen frontier models. An 18-model family (0.8B–1T+ parameters), co-developed with Zoo Labs Foundation, with owned weights and full provenance.
Structure
Each paper lives in its own subdirectory:
papers/
├── shared/ # cover styles, lstlang.tex, paperkit
│ ├── hanzocover.sty
│ └── lstlang.tex
├── <paper-slug>/
│ ├── <paper-slug>.tex # main file (\input's sections)
│ ├── <paper-slug>.pdf # compiled output
│ └── sections/ # modular sections
│ ├── 01-intro.tex
│ ├── 02-architecture.tex
│ ├── 03-protocol.tex
│ ├── ...
│ └── 99-bibliography.tex
└── INDEX.md # auto-generated catalogue
Building
cd <paper-slug>
TEXINPUTS=".:..:" latexmk -pdf <paper-slug>.tex
Or build all — every paper, and a nonzero status naming any that did not compile:
make -k all
Index
See INDEX.md for full catalogue of papers.
Conventions
- One paper, one directory. No top-level .tex files.
- Modular sections. Main .tex \input's
sections/NN-name.texfiles for easy editing. - Shared cover. All papers use
\usepackage{shared/hanzocover}and\hanzocoverpage. - Shared lstlang. All papers use
\input{shared/lstlang}after\usepackage{listings}. - No AI slop. Technical, dense, citation-supported.
- One paper per concept. Updates over time via versioning, not duplication.