# Make science agentic. — Hanzo AI

> Hanzo Science is the universal research operating system. Design new AI architectures, discover materials, investigate biology, simulate physical systems, and explore the universe.

Hanzo Science · Universal Research Operating System

# Make science agentic.

The AI platform for research, discovery, and engineering.

Turn questions into investigations, ideas into experiments, and experimental results into the next breakthrough. Hanzo Science brings AI agents, scientific models, data, computation, and research tools together in one environment.

- Design new AI architectures · Discover materials · Investigate biology · Simulate physical systems · Explore the universe
- One platform. Every discipline. Research without boundaries—from individual researchers to the world's largest R&D organizations.

[Explore Hanzo Science](#introducing-hanzo-science)[See research workflows](#how-it-works)[Partner with us](https://hanzo.ai/contact-sales)

01Universal R&D Loop

02Hanzo AI Lab

03Research Graph & Control Plane

04Interactive Workbenches

057-Dimension Verification

06Commercial Tiers & Access

Hanzo Science Workspace · Select a research field to see how the same product adapts

prog_ai_kai_latency_v4

AI & Machine Learning

Biology & Drug Discovery

Aerospace & Astrophysics

Materials & Energy

Semiconductors & EDA

Home· Programs & Graph· Experiments· Evidence & Data· Optimization· Review· Compute· Governance

AI & Machine Learning04 · Evaluate → 05 · Optimize

Reduce high-cardinality multiple-choice inference latency in Kai by 35% while preserving numerical equivalence, calibration (ECE <= 0.012), and stability.Enso (orchestration) · Zen (reasoning) · Kai (policy gate)Hanzo ML / Engine

Ray Tune

MLflow Registry

Calibration Harness

p99 latency-38.4% (14.2 ms)

Calibration ECE0.011 (target <= 0.012)

Hardware & seed8× H100 · seed=2026

kai_routing_sweep.parquetDataset · sha256:9a4c…e102

latency_frontier_fig2.pdfFigure v3 · sha256:3d18…88b4

kai_release_qual.ro-crate.jsonRO-Crate · sha256:7f01…c419

7/7 verification dimensions passed · Approved by @QA + Systems Lead

01One scientific execution engine across every research-driven industry

Scientific and engineering breakthroughs are slowed by the gaps between literature, data, code, simulators, instruments, compute clusters, and peer review. Hanzo Science unifies the six-stage research loop so every hypothesis, simulation, figure, and decision stays linked to the exact code, container digest, and evidence that produced it.

The 6-stage universal research loop — backed by a branching Research GraphBranch competing hypotheses · Run in parallel · Preserve negative results

01 · DiscoverReview knowledge, identify gaps & generate ideasIngest papers, patents, ELN/LIMS entries, astronomical catalogs, and prior negative findings with direct DOI/PMID/arXiv provenance.

02 · DesignForm competing hypotheses, select methods & planBranch competing hypotheses in a research graph, define baselines, controls, simulation matrices, and pre-registered acceptance thresholds.

03 · ExecuteRun code, train models, simulate, test or measureDispatch reproducible containers, live Hanzo Notebook kernels, OpenMDAO / Phonopy solvers, targets + crew DAGs, and Slurm/cloud GPU sweeps.

04 · EvaluateAnalyze evidence, uncertainty, failures & resultsAudit statistical validity, conservation laws, numerical equivalence, cycle hysteresis, calibration error, and assertion-level source support.

05 · OptimizeCompare alternatives, refine methods & allocate budgetMaximize validated information gained per unit of compute using Bayesian optimization, active learning, and Pareto frontier search.

06 · CommunicateReproduce findings, review, publish or transferCompile LaTeX manuscripts, CDISC ADaM/rtables TLFs, qualified model cards, and RO-Crate / W3C PROV-O lineage bundles.

Supports both fundamental discovery and applied R&DFundamental DiscoveryExploratory researchGenerate new questions, develop theories, investigate unexpected observations, and branch into competing explanations when there is no predetermined optimization target.Example: What new attention or state-space architectures might overcome long-context efficiency limitations, or which cellular mechanisms drive resistance in a cohort?First-class support for mathematical proofs (Lean 4), theoretical modeling, systematic literature synthesis, observational studies, and exploratory simulations.

Applied R&D & EngineeringGoal-directed researchOptimize a defined outcome, compare candidate designs, test a specific hypothesis, and solve a concrete engineering or clinical problem under resource bounds.Example: Reduce Kai high-cardinality inference latency by 35% while preserving numerical equivalence and calibration, or maximize rocket nozzle Isp within wall thermal limits.First-class support for controlled ablations, multi-objective Pareto sweeps, OpenMDAO multidisciplinary design optimization, and release qualification gates.

Three markets served simultaneously — one research engine, every scale of discovery01 · Enterprise R&DEnterprise R&D OrganizationsAI laboratories, pharmaceutical, semiconductor, aerospace, energy, materials, automotive, and industrial companies.Unify fragmented tools, private data lakes, HPC clusters, and cross-functional review gates under one governed research operating system.

02 · Research InstitutionsUniversities, National Labs & HospitalsAcademic laboratories, national user facilities, observatories, research hospitals, and government science agencies.Connect shared supercomputers, telescope/beamline pipelines, multi-institution datasets, and FAIR / RO-Crate reproducibility.

03 · Individual Scientists & StartupsIndependent Researchers & Deep-Tech StartupsPrincipal investigators, computational scientists, startup founders, and small R&D teams.Spin up specialist scientific agents, live notebooks, domain simulators, and elastic GPU compute without enterprise DevOps overhead.

One platform for the full spectrum of research across 20 disciplinesBasic science · Applied research · Product engineering · Commercial R&D

Artificial intelligenceNew model architectures, training methods, agent systems, scaling laws, evaluations

Computer scienceAlgorithms, compilers, distributed systems, programming languages, formal verification

MathematicsConjectures, Lean 4 formal proofs, symbolic computation, numerical analysis

Theoretical physicsMathematical modeling, lattice/field simulations, theory comparison, conservation checks

Astrophysics & cosmologyAstronomical data analysis, stellar modeling, cosmological N-body rollouts, Bayesian inference

Aerospace & astronauticsAerodynamics, orbital analysis, thermal modeling, structural FEA, OpenMDAO design optimization

Mechanical engineeringComputational fluid dynamics (CFD), finite-element analysis, thermal systems, structural design

Electrical engineeringCircuit design, SPICE simulation, RF/signal processing, feedback control systems

SemiconductorsRTL microarchitecture exploration, device TCAD simulation, OpenROAD STA, formal SVA verification

RoboticsEmbodied AI, physics simulation, sensor fusion, trajectory control, reinforcement learning

Materials scienceCrystal structures, electronic properties, 8×8×8 q-mesh DFPT phonon stability, candidate screening

ChemistryMolecular properties, reaction modeling, homogeneous/heterogeneous catalysts, quantum chemistry

BiologyGenomics, proteomics, single-cell atlases (Scanpy/scVI), cellular systems, computational evolution

Drug discoveryTarget investigation, LillyMol iwdemerit/trxn SAR screening, OpenMM ΔΔG, experimental planning

Medicine & clinical researchLongitudinal MMRM (mmrm/brms.mmrm), trial simulation (rfacts), CDISC ADaM (admiral/rtables)

NeuroscienceNeural spike/field recordings, connectome modeling, computational neuroscience, biological networks

EnergySolid-state batteries, fusion plasma equilibrium, photovoltaic materials, power-grid optimization

Earth & planetary sciencesGeophysics, planetary geology, atmospheric dynamics, ERA5/CMIP6 spectral climate models

AgriculturePlant biology, quantitative genetics, crop trait selection, agro-environmental modeling

Industrial R&DManufacturing processes, reliability engineering, physics-informed digital twins, design of experiments

Interactive Figure, Manuscript & Code WorkbenchInspect figures, click directly on plot annotations to refine the underlying Python/R script, switch between physics, fluid PDE, chemistry, Lean 4 proof, and clinical R&D artifacts, and audit the execution log.

Hanzo ScienceBeta

Cross-species Samosa

New

Customize

Files

Today◦Samosa Manuscript

◦Samosa Editorial Review

◦Samosa Model Training

◦Samosa Benchmarking

hanzo/z

samosa_fig1a.png

samosa_fig1a.png

a138 species · 5,672 cell types · one shared embeddingv2

1

1

Send

cell-type family●neuron

●muscle

●immune

●ciliated

●germline

●stem/progenitor

i ciliated cells27 types · 18 species · 6 phyla• Ctenophore comb-plate cells• Sponge choanoblasts• Human airway multiciliated• Schistosome flame cells• Xenoturbella ciliated

ii striated muscle116 types · 40 species · 7 phyla• Acoel muscle• C. elegans body-wall muscle• Ctenophore striated muscle• Cnidarian muscle• Zebrafish fast muscle

iii immune84 types · 34 species · 6 phyla• Arrow worm hemocytes• Cow myeloid• Oyster hemocytes• Filarial nematode coelomocyte

samosa_fig1a.png

v2

Open Hanzo Notebook

Code

Execution Log

Messages

Environment

Review ✓

Download script

Inputsfig4_atlas_callouts.csvfig4_atlas_centroids_m138ea.csv

54
55 apply_nature_style()
56 mpl.rcParams['savefig.bbox'] = None
57 mpl.rcParams['font.sans-serif'] = ['Arial']
58 mpl.rcParams['font.family'] = 'sans-serif'
59
60 centroids = pd.read_csv("fig4_atlas_centroids_m138ea.csv")
61 boxes_df = pd.read_csv("fig4_atlas_inset_boxes.csv")
62 callouts = pd.read_csv("fig4_atlas_callouts.csv")
63
64 HERO = {"neuron":"steelblue","muscle":"olive","immune":"forestgreen",
65 "ciliated":"darkturquoise","germline":"orchid","stem/progenitor":"peru"}
66 PHYL_PAL2 = {"Ctenophora":"saddlebrown","Porifera":"slateblue","Placozoa":"darkslategray",
67 "Xenacoelomorpha":"cadetblue","Chaetognatha":"darkslateblue",
68 "Ambulacraria":"teal","Cephalochordata":"dimgray"}
69 TAG_ROMAN = {'a':'i','b':'ii','c':'iii'}
70 INSET_NAMES = {'a':'ciliated cells','b':'striated muscle','c':'immune'}
71 ITALIC_SPLIT = {'Xenoturbella ciliated':('Xenoturbella',' ciliated'),
72 'C. elegans body-wall muscle':('C. elegans',' body-wall muscle')}
73 TAG_COL = 'black'
74 halo = [pe.withStroke(linewidth=2.2, foreground='white')]
75 halo_thin = [pe.withStroke(linewidth=2.0, foreground='white')]
76
77 inset_info = {}
78 for _, row in boxes_df.iterrows():
79 tag = row.tag; fam = row.family; x0, y0, w, h = row.x0, row.y0, row.w, row.h
80 inbox = centroids[(centroids.umap_x >= x0) & (centroids.umap_x <= x0 + w)]
81 target = inbox[inbox.family == fam]
82 tx0, tx1 = target.umap_x.min(), target.umap_x.max(); ty0, ty1 = target.umap_y.min(), target.umap_y.max()
83 mx, my = 0.05 * (tx1 - tx0), 0.05 * (ty1 - ty0)
84 inset_info[tag] = dict(fam=fam, xlim=(tx0 - mx, tx1 + mx), ylim=(ty0 - my, ty1 + my),
85 n_types=row.n_types, n_species=row.n_species)

02Built inside a real AI research lab — Hanzo AI runs on Hanzo Science

Hanzo AI is the first full-scale research organization operating on Hanzo Science. Our teams use the same ResearchProgram graph, evaluation harnesses, and 7-dimension verification engine to design architectures, curate pretraining mixtures, quantize KV caches, optimize Kai decision latency, and qualify Enso and Zen model releases.

First-class AI model research lifecycle — integrated with MLflow, Ray Tune & Hanzo MLBranch, compare, rerun, or return to any earlier research stage

01 · DesignArchitecture

02 · DataDatasets

03 · TrainOptimization

04 · TuneAdaptation

05 · EvaluateBenchmarks

06 · ProfileEfficiency

07 · QualifyRelease gates

08 · DeployControlled rollout

12 AI research workloads coordinated across the Hanzo laboratoryWhere Hanzo invents, trains, optimizes & qualifies its own technology

Model architecturesInvestigate attention mechanisms, model layers, routing structures, memory, multimodal architectures, and alternative designs.

PretrainingDataset preparation, tokenization, scaling experiments, distributed training, checkpoints, and convergence analysis.

Fine-tuningSFT, LoRA/QLoRA, preference optimization, distillation, reinforcement learning, and hyperparameter searches.

Model optimizationQuantization, pruning, sparsity, speculative decoding, compilation, and model compression.

Inference enginesKernel benchmarking, KV-cache optimization, batching, scheduling, graph execution, and serving performance.

Agentic harnessesTool use, planning, memory, context management, sandbox execution, task completion, recovery, and multi-agent coordination.

ML frameworksTraining runtimes, tensor operations, device backends, distributed execution, and numerical correctness.

Model evaluationAccuracy, reasoning, calibration, stability, safety, regression, latency, throughput, and cost.

Data researchCorpus discovery, quality filtering, synthetic data, labeling, contamination checks, and dataset versioning.

System optimizationGPU/CPU profiling, allocation, networking, storage, compiler performance, and infrastructure economics.

Automated researchHypothesis generation, controlled experiments, ablations, statistical comparisons, and next-experiment recommendations.

Release qualificationBenchmark evidence, reproducibility, release gates, regression analysis, model cards, and rollback readiness.

Walkthrough · Science improves Hanzo's own inference architectureObjective: Reduce Kai's high-cardinality multiple-choice inference latency while preserving numerical equivalence, accuracy, and stability

1. Investigate the bottleneckRead profiling traces, relevant source code, existing benchmarks, hardware characteristics, and previous experimental findings.

2. Form competing hypothesesEvaluate microbatching, caching, vectorization, memory-layout changes, FP8 KV-cache quantization, and speculative tree routing.

3. Generate experimentsCreate isolated branches, test implementations, reference comparisons, and reproducible benchmark configurations.

4. Execute and measureUse the coding harness and available GPU compute to build, test, profile, and measure candidate changes under equivalent conditions.

5. Evaluate statistically & numericallyCompare p50/p95/p99 latency distributions, throughput, memory usage, numerical equivalence, calibration (ECE), and regressions.

6. Recommend a qualified changePresent the best supported result, code diff, test results, risks, and release requirements. Request approval before merging or deploying.

Final artifacts: Experiment report · Benchmark dataset · Source-code diff · Execution logs · Reproducible container · Release recommendation

Mapped directly to Hanzo's existing stackHanzo Science is the common research orchestration and evidence layer connecting training, code harnesses, inference engines, and team approvals.

ML framework

New operators, training algorithms, optimizers, kernels, device support

Engine

Model execution, inference scheduling, batching, memory, performance

Node / compute

Distributed workloads, hardware topology, accelerators, profiling

Cloud

Resource scheduling, GPU allocation, cost, cluster experiments

Sandbox

Isolated builds, test environments, scientific code execution

Agents / Dev harness

Autonomous experiments, code changes, test runs, recovery

Enso

Model routing research, selection policies, orchestration

Kai

Structured decision research, policy evaluation, qualification

Zen

Generative-model research, fine-tuning, reasoning evaluations

Gateway / serving

API performance, routing quality, latency, production regression

Hanzo Team

Collaboration, scientific approvals, research program coordination

03Universal Research Graph, Control Plane & Open Standards

Instead of building separate silos for each discipline, Hanzo Science represents every investigation as a typed ResearchProgram / ScienceWorkload object orchestrated by Enso, reasoned over by Zen, and governed by Kai—with native export to CWL, RO-Crate, and W3C PROV-O.

Science DiscoveryIdeas, evidence, literature & hypothesesIngest primary literature, patents, datasets, and anomalies; formulate branching questions and competing hypotheses.

Science ExperimentsTraining, simulation, computation & testsExecute controlled experiments across Hanzo ML, Engine, Sandboxes, OpenMDAO, Phonopy, LillyMol, OpenMM, and Lean 4.

Science OptimizationSearch, comparisons & resource allocationMaximize Research Value = (validated information or progress gained) / (resources consumed) via Bayesian and Pareto search.

Science KnowledgeFindings, artifacts, provenance & collaborationLink every finding (including negative results) to its exact run, dataset, environment, and RO-Crate / W3C PROV-O record.

Universal Research Graph — Canonical ResearchProgram ModelRO-Crate · CWL · W3C PROV-O

A finding is traceable to a specific analysis, run, dataset, method, environment, and hypothesis. Later results can contradict earlier findings without erasing historical provenance.interface ResearchProgram {
id: string;
organization_id: string;
domain: string;
title: string;

questions: ResearchQuestion[];
hypotheses: Hypothesis[];
methods: ResearchMethod[];

evidence: EvidenceRecord[];
experiments: Experiment[];
runs: ExecutionRun[];

observations: Observation[];
findings: ResearchFinding[];
decisions: ResearchDecision[];

artifacts: ArtifactReference[];
provenance: ProvenanceGraph; // RO-Crate + W3C PROV-O

compute_budget: ResourceBudget;
approval_policy: ApprovalPolicy;
collaborators: ResearchCollaborator[];

status: ResearchStatus;
}

Deep integration with the Hanzo agentic harness@SciResearch planning, scientific investigation, hypothesis tracking

@DevSource-code changes, instrumented test programs, computational pipelines

@DataDataset preparation, data analysis, statistics, lineage

@QAVerification, regression testing, reproducibility checks

@CyberEnvironment isolation, credential boundaries, tool security

@DocResearch documentation, manuscripts, and technical reports

@ManaProject coordination, compute budgets, milestones, and review queues

Unified Research Command Interface (Control Plane API)REST & MCP

POST /v1/science/programs
POST /v1/science/questions
POST /v1/science/hypotheses
POST /v1/science/experiments
POST /v1/science/experiments/{id}/runs
GET /v1/science/runs/{id}/metrics
POST /v1/science/experiments/{id}/compare
POST /v1/science/decisions
GET /v1/science/artifacts/{id}/provenance
POST /v1/science/reviews

04How it works across disciplines

Step through multi-discipline simulation and formal verification pipelines, dispatch GPU parameter sweeps alongside a shared live Hanzo Notebook kernel, and audit primary literature against NCBI E-utilities and Crossref before compiling LaTeX manuscripts.

Interactive walkthroughs across disciplines

AI & ML Systems

Physics & Materials

Semiconductors & EDA

Earth & Fluid PDEs

Chemistry & MedChem

Mathematics

Genomics & Single-Cell

Clinical & Proteomics R&D

Crystal relaxation & DFPT phonon stability48 symmetries · 0.01 Å

01. Build supercells

02. 8×8×8 q-mesh & ASR

03. Verify DFPT tolerance

Pause

1x

00:00 / 01:20

Screen candidate solid-state materials, evaluate forces on an 8×8×8 q-mesh, and verify acoustic sum rule (ASR) stability before synthesis.01. Build supercellsLoad candidate crystal structures and generate finite-displacement 2×2×2 supercells.

02. 8×8×8 q-mesh & ASREvaluate atomic forces on GPU, enforce translational acoustic sum rule at Γ, and solve dynamical matrix.

03. Verify DFPT toleranceReviewer confirms zero imaginary modes beyond acoustic sum rule tolerance at Γ and locks the figure.

Step 01 script

atoms = read("candidate_lattice.cif")
phonon = Phonopy(atoms, supercell_matrix=[[2,0,0],[0,2,0],[0,0,2]])
phonon.generate_displacements(distance=0.01)

Manages cluster compute and shares a live Hanzo Notebook kernelDispatch multi-arm GPU sweeps with full software version, hardware, seed, and baseline tracking while inspecting live variables in memory.

Hanzo ScienceBeta

scRNA-seq

New

Customize

Files

Active◦Samosa Manuscript

◦Cross-species scRNA-s...

◦SCVI Hyperparameter S...

8

◦Phonon & PDE Solver

3

hanzo/z

Scvi hyperparameter screenDispatching the 8-arm scVI sweep to lab_cluster A100s — n_latent ∈ {10, 20, 30, 50} × n_layers ∈ {1, 2}, 40k cells × 2,000 HVGs, batch_key="sample_id", 50 epochs, seed 0.Env: scvi-tools 1.1.2 · scanpy 1.10.1 · PyTorch 2.3.0+cu121 · 8× A100-80GB (2.08 GPU-hrs)Baseline (unintegrated PCA / Harmony): cLISI 0.89, iLISI 0.61 → Arm #6 (d=30, L=2): ELBO -412.8, cLISI 0.94, iLISI 0.88

armn_latentn_layerslabel

1101d=10 L=1 · scVI COVID-PBMC (1/8)

2102d=10 L=2 · scVI COVID-PBMC (2/8)

3201d=20 L=1 · scVI COVID-PBMC (3/8)

4202d=20 L=2 · scVI COVID-PBMC (4/8)

5301d=30 L=1 · scVI COVID-PBMC (5/8)

6302d=30 L=2 · scVI COVID-PBMC (6/8)

7501d=50 L=1 · scVI COVID-PBMC (7/8)

8502d=50 L=2 · scVI COVID-PBMC (8/8)

Stephenson 2021 COVID PBMC — 40k cells, 2,000 batch-aware HVGs, no integrationBy sequencing site

By author cell type

Remote · 8lab_cluster · d=10 L=1 · scVI CO...

16m 2s

lab_cluster · d=10 L=2 · scVI CO...

15m 42s

lab_cluster · d=20 L=1 · scVI CO...

15m 19s

lab_cluster · d=20 L=2 · scVI CO...

14m 58s

lab_cluster · d=30 L=1 · scVI CO...

14m 36s

lab_cluster · d=30 L=2 · scVI CO...

14m 16s

8 running · 16m 2s

Notebookliver-pipelineShared with the agent

● Live · python

Cell

[28]python

Run

import pandas as pd
pd.set_option('mode.string_storage', 'python')
import numpy as np, scanpy as sc, anndata as ad, scipy.sparse as sp

a = sc.read_h5ad("covid_pbmc_40k_hvg.h5ad")
print("loaded:", a.shape, "uns keys:", list(a.uns.keys()))

# minimal, version-portable object: counts + obs + var only
keep_obs = ["sample_id", "donor_id", "Site", "initial_clustering",
"author_cell_type", "disease", "Status",
"Status_on_day_collection_summary", "cell_type"]
keep_obs = [c for c in keep_obs if c in a.obs.columns]
obs = a.obs[keep_obs].copy()
for c in obs.columns:
obs[c] = pd.Index(np.asarray(obs[c], dtype=object))
obs.index = pd.Index(np.asarray(obs.index, dtype=object))
var = pd.DataFrame(index=pd.Index(np.asarray(a.var_names, dtype=object)))

clean = ad.AnnData(
X=sp.csr_matrix(a.layers["counts"]),
obs=obs, var=var,
)
clean.layers["counts"] = clean.X.copy()
clean.write_h5ad("covid_pbmc_40k_hvg.h5ad")
print(f"rewritten: {os.path.getsize('covid_pbmc_40k_hvg.h5ad')/1e6:.1f} MB")

> outputwrote covid_pbmc_40k_hvg.h5ad

Python — liver-pipeline kernelShared state active

>>> clean.shape(40000, 2000)

>>> list(clean.obs.columns[:4])['sample_id', 'donor_id', 'Site', 'initial_clustering']

Eval

Independent database retrieval & assertion-level citation reviewRather than trusting unverified model prose, the Reviewer agent queries NCBI E-utilities, Crossref, OpenAlex, and arXiv directly and verifies claim-level table support before compiling the PDF.

Hanzo ScienceBeta

Literature review

New

Customize

Files

Active◦Samosa Manuscript

◦Cross-species scRNA-s...

◦SCVI Hyperparameter S...

8

◦Phonon & PDE Solver

3

hanzo/z

Cross-species scRNA-seq Integration...Write a literature review on cross-species single-cell RNA-seq integration. Pull the primary methods papers and recent benchmarks. Output the report as a LaTeX doc and a compiled PDF.

Ran 4 searches, loaded 2 skills, managed environments, +2 more

10 steps

Dispatching five parallel literature-retrieval tracks — PubMed primary methods, bioRxiv preprints, OpenAlex citation counts, CELLxGENE multi-species atlas inventory, and orthology-free embedding methods.Dispatching PubMed bioRxiv OpenAlex CELLxGENE sub-agents

142 lines of output

Reviewer · 1 finding

WarnPMID 31178118 assigned to both LIGER and Seurat v3 integration in the plan

In the generate_plan PubMed delegation step the agent writes “LIGER (31178118), Seurat v3 integration (31178118)” — the same PMID for two distinct primary methods papers. The same plan’s OpenAlex step assigns them different DOIs (Seurat v3 10.1016/j.cell.2019.05.031, LIGER 10.1016/j.cell.2019.05.006), so the plan is internally inconsistent and at least one PMID is wrong. None of the PMIDs or DOIs in msg[10] trace to any in-window tool output — no PubMed/CrossRef/OpenAlex lookup has run (exec-log rows 04449b53, 57baea0a, 254f1378 are the two skill kernel.py auto-loads and the env-filter bash; msg[4] tool_results are skill-catalog listings only).The agent reads these findings and self-corrects in its next message.

exec-log 8f91c204 · Independent NCBI E-utilities & Crossref verificationRetrieved primary records and verified assertion-level support before LaTeX compilation: Stuart et al. (Seurat v3) matched to PMID 31178118 / DOI 10.1016/j.cell.2019.05.031; Welch et al. (LIGER) matched to PMID 31178122 / DOI 10.1016/j.cell.2019.05.006.

⇄ all 5 agents done✓ Reviewing

review.pdf

Open Hanzo Notebook

Cross-species single-cell RNA-seq integration: from one-to-one orthologs to protein-language-model embeddingsLiterature synthesis compiled from PubMed, bioRxiv, OpenAlex and CELLxGENE Discover

Papers24

Years2018–2025

Methods15

Species pairs6 benchmarked

Top-citedSeurat v3 (16,859 cit.)

Most recentTranscriptFormer (2025)

Figure 1: Fifteen integration methods (2018–2024) coloured by orthology strategy; OpenAlex citation counts on log scale.

1 Problem statementComparative single-cell transcriptomics asks whether a cell type in one species has a homolog in another, and how its expression program has been conserved or rewired. The technical obstacle is that any two species’ transcriptomes live in different gene coordinate systems.

L(θ, φ) = E_q[log p_θ(x | z, s)] - D_KL(q_φ(z | x, s) || p(z)) (1)M_ij = ⟨x_i, H y_j⟩ / (||x_i|| ||H y_j||) (2)

SAMap [10] replaced the hard ortholog filter with a soft, BLAST-derived gene-homology weight matrix H and iteratively refined it against cell-level manifold alignment. SATURN [11] and UCE [12] replace H entirely with ESM2 protein-language-model embeddings in a shared 1,280-dimensional protein space.2 Ortholog-subsetting methodsSeurat v3 [1] (PMID 31178118) finds canonical-correlation vectors u, v maximising corr(u^T X, v^T Y) over shared-ortholog matrices, while LIGER [2] (PMID 31178122) factorises each dataset via integrative NMF X_k ≈ (W + V_k) H_k.

Domain packs, scientific databases, simulators & compute connectors

All

AI & ML

Semiconductors

Physics & Materials

Earth & Climate

Chemistry

Mathematics

Biology & Clinical

Compute & Standards

Hanzo Enso, Zen, Kai & MLAI & ML SystemsImplemented

Multi-agent orchestration (Enso), scientific reasoning (Zen), typed decision latency sweeps (Kai), and distributed PyTorch/vLLM harnesses.

OpenROAD, Verilator & SymbiYosysSemiconductors & EDATested

SystemVerilog RTL elaboration, static timing analysis (STA), PPA Pareto sweeps, SPICE waveforms, and formal SVA verification.

Materials Project, Phonopy & MACEPhysics & MaterialsImplemented

Crystal structures, DFT formation energies, 8×8×8 q-mesh DFPT phonon dispersion with ASR enforcement, and machine-learned force fields.

CERN Open Data & INSPIREPhysics & MaterialsTested

High-energy physics datasets, ROOT trees, kinematics histograms, and particle physics literature synthesis.

ERA5, CMIP6 & EarthDataEarth & ClimateImplemented

Global atmospheric reanalysis Zarr tensors, ocean circulation grids, and spectral PDE conservation audits.

LillyMol, RDKit & OpenMMChemistry & MedChemImplemented

iwdemerit 275-rule MedChem demerit filter, tsubstructure SMARTS, trxn reaction enumeration, dicer cleavage, and OpenMM ΔΔG.

Lean 4, Mathlib & SymPyMathematicsImplemented

Formal theorem verification in Lean 4 kernel, tactic search, symbolic algebra, and counterexample checks.

CELLxGENE, Scanpy & scVIGenomics & Single-CellImplemented

Multi-species cell atlases, AnnData h5ad slices, and variational manifold alignment benchmarked against PCA/Harmony.

mmrm, brms.mmrm, rfacts & TSDTClinical BiostatisticsImplemented

Frequentist & Bayesian longitudinal MMRM, adaptive clinical trial simulation (rfacts), and subgroup detection trees (TSDT).

pharmaverse (admiral, rtables, Tplyr)CDISC & RegulatoryImplemented

Automated CDISC SDTM/ADaM dataset derivation (ADSL, ADAE, ADLB, ADTTE) and regulatory TLF tables.

BioRels, surfaltr & ProteomicsProteomics & TargetsTested

Longitudinal clinical proteomics differential abundance, transmembrane surfaceome topology (surfaltr), and BioRels schemas.

UniProt, RCSB PDB & EnsemblStructure & SequenceImplemented

Experimental 3D structures, co-folding confidence maps, and variant annotations.

targets, crew & pytest-wdlPipelines & WDLImplemented

Reproducible R/Python targets pipeline DAGs, crew distributed Slurm/K8s workers, and pytest-wdl workflow verification.

CWL, RO-Crate & W3C PROV-OProvenance & StandardsImplemented

Portable Common Workflow Language pipelines, Research Object Crates (RO-Crate), and W3C PROV-O lineage export.

arXiv, OpenAlex, Crossref & PubMedLiterature & CitationsImplemented

Direct NCBI E-utilities, Crossref, OpenAlex, and arXiv retrieval with independent assertion-level citation verification.

Hanzo Notebook & SSH / Slurm ClustersLive ComputeImplemented

Shared Python/Julia/R/SQL kernel + scaling from local workstation to Slurm, Kubernetes, and cloud GPU nodes.

05Scientific verification, graduated autonomy & flagship workloads

High-consequence R&D cannot rely on opaque summaries. Hanzo Science evaluates every workload across seven independent verification dimensions and four graduated autonomy modes.

7-dimension scientific verification — never a single unexamined green checkmark01 · Source integrityDoes the cited source exist and resolve?Independent NCBI E-utilities, Crossref, OpenAlex, arXiv, and patent registry lookup verifying DOI, PMID, publication year, and retraction status.

02 · Claim supportDoes the cited passage or table actually support the assertion?Assertion-level extraction comparing the manuscript sentence against the primary source table, sample size, and experimental conditions.

03 · Computational integrityDid the code execute cleanly and produce the reported numbers?Direct kernel execution check confirming zero silent NaNs, non-zero exit codes, or mismatches between figure annotations and output tensors.

04 · ReproducibilityCan another scientist re-run the exact computation?Records OCI container digest, conda/lockfile hashes, random seeds, input dataset SHA-256, hardware topology, and RO-Crate bundle.

05 · Method validityAre controls, baselines, units, and physical laws respected?Domain-specific checks: acoustic sum rule (ASR) at Γ for phonons, cycle closure hysteresis for FEP, PCA/Harmony baselines for scVI, and STA corners for RTL.

06 · Scientific uncertaintyAre error bars, sensitivity bounds, and alternative hypotheses reported?Requires confidence/credible intervals, convergence diagnostics (Rhat, ESS), and sensitivity sweeps across key hyperparameters.

07 · Domain reviewWhich decisions require accountable human sign-off?Routes synthesis nominations, clinical TLF locks, silicon tapeout checklists, and model releases to named human reviewers via Kai policy gates.

4 graduated autonomy modes governed by Kai policy gatesMode 1 · ObservePassive monitoring & provenance captureIndexes literature, notebooks, cluster jobs, and instrument outputs; flags anomalies or citation mismatches without mutating files.

Mode 2 · RecommendHuman-gated experimental proposalsDrafts code patches, figure refinements, next parameter arms, and synthesis candidates; waits for explicit scientist approval.

Mode 3 · Execute approvedBounded execution within pre-approved envelopesRuns approved CWL/WDL/targets workflows and GPU sweeps within strict compute budgets, connector allowlists, and container sandboxes.

Mode 4 · Supervised automationClosed-loop active learning with human checkpointsIterates design-execute-evaluate cycles autonomously across simulation or assay rounds, pausing at Kai policy gates for material decisions.

Phased architecture rollout — validated in Hanzo AI Lab firstP0 · Foundation & AI LabUnified research programs, Hanzo AI Lab integration, reproducible comparisons & checkpoint lineageResearch survives session restarts; runs real model, inference-engine, and harness experiments traced to exact inputs and versions.

P1 · Domain Packs & OptimizationAgentic experiment planning, Biology/Materials/Physics/Engineering domain packs & publication reviewsGenerates competing hypotheses, executes the same workload lifecycle across domain solvers (OpenMDAO, Phonopy, LillyMol), and produces defensible reports.

P2 · Continuous & Cross-Org R&DEvent-driven continuous research & federated multi-institution R&D coordinationRe-evaluates approved investigations when new evidence arrives and coordinates governed collaboration across laboratories and institutions.

Measure research productivity, not raw AI activityThe goal is to increase the rate at which researchers generate reliable, validated knowledge—including well-designed negative results that rule out dead ends.

Time from research hypothesis to a reproducible experiment

Valid experiments completed per unit of compute and researcher effort

Model or engineering improvements that survive independent held-out evaluation

Efficiency gains in latency, throughput, memory, and infrastructure cost

Experiment reproducibility rate, regression detection, and preserved negative results

Time from accepted research finding to qualified release or protocol transfer

Commercial editions — priced for lower cost per validated scientific resultScience IndividualIndependent researchers, postdocs & studentsFull Science Workbench, shared Hanzo Notebook kernel, literature & citation verifier, RO-Crate export, and local/SSH compute attach.

Science TeamAcademic labs, biotech/hardware startups & R&D groupsShared ResearchProgram graphs, hanzo.team multi-agent channels (@Sci, @Data, @Dev), Slurm/Kubernetes crew dispatch, and 7-dimension review.

Science EnterpriseGlobal pharmaceutical, semiconductor, aerospace, materials & industrial R&DSSO/SAML, RBAC, connector allowlists, Kai governance gates, audit logs, GxP/CDISC, OpenMDAO & EDA domain packs, and dedicated cluster integration.

Science PrivateNational labs, defense, sovereign & air-gapped facilitiesCustomer-managed VPC or on-premise supercomputer deployment, zero external data egress, local Zen/Enso/Kai weights, and custom data residency.

Four flagship domain workloads running on Hanzo Science01 · AI & Machine LearningModel architecture, quantization & Kai latency qualificationResearch teams coordinate pretraining ablations, KV-cache quantization, and speculative decoding sweeps while auditing calibration error, regressions, and release criteria.

Enso + Zen + Kai · Reproducible evaluation harnesses

02 · Life Sciences & PharmaTarget discovery, LillyMol SAR & longitudinal clinical biostatisticsMedicinal chemistry and biostatistics groups connect 275-rule iwdemerit filtering and alchemical ΔΔG cycles with WDL proteomics and CDISC ADaM regulatory tables.

LillyMol + OpenMM + targets/crew + brms.mmrm

03 · Semiconductors & EngineeringMicroarchitecture PPA exploration & formal SVA verificationHardware architects evaluate RTL parameter sweeps across process-voltage-temperature corners and verify credit-return liveness before physical synthesis.

Verilator + OpenROAD STA + SymbiYosys proofs

04 · Materials & EnergySolid-state electrolyte screening & DFPT phonon stabilityMaterials laboratories rank candidate lattices against formation energy, ionic migration barriers, and acoustic sum rule dynamical stability prior to synthesis.

8×8×8 q-mesh + ASR at Γ + NEB diffusion barriers

Scientific review often turns on a small detail. Suppose a veterinary study reports that missed appointments fell from 12% to 9% after clinics introduced reminders. That’s a drop of three percentage points, or 25% relative to the original rate. Both figures are correct. Whether reminders caused the decline is a separate question, and the answer depends on how the study was designed.

Hanzo Science helps researchers keep those distinctions visible as they work.

## Read across the literature and your own records

Start with a question, a collection of papers, or an existing dataset. Hanzo searches the public web alongside the articles, notes, and datasets your team adds. It can find an exact phrase or related work that uses different terminology, then cite the material behind its answer.

For a team studying appointment reminders, that could mean finding papers that describe missed visits as nonattendance, pulling out how each study defined a missed visit, and identifying which studies used a comparison group. A colleague can open the cited source and check whether the studies belong in the same analysis.

## Run the analysis and keep the evidence

Use a Hanzo sandbox to run a calculation, or a GPU machine for work that needs more compute. Keep the output alongside the question it answers. If two clinics counted cancellations differently, a researcher can rerun the comparison using a common definition and see how much the finding changes.

Hanzo also records when evidence entered the research record. When a new paper changes the team’s conclusion, researchers can revisit what they knew at the time of the original decision and identify the evidence that changed their minds.

## Test the research method itself

Teams can use Hanzo Evals to compare prompts or models against the same set of expert-reviewed questions. A lab might score three things separately: whether the system found the right paper, whether the cited passage supports the claim, and whether the calculation is correct. That makes it easier to see exactly where a revised method improves the work and where it still needs expert judgment.

## Get started

Bring a question your team has already answered, the sources you trusted, and the result you accepted. Run that question through Hanzo Science and have a colleague review the evidence. A useful first test is whether that colleague can explain how the team reached its conclusion and reproduce the calculation behind it.

[Explore Hanzo for research](https://hanzo.ai/solutions/research)

06Getting started with Hanzo Science

macOS, Linux, Web & SSH / Slurm

Available in beta across Individual, Team, Enterprise, and Private deployments. Run locally or attach to your own Linux workstation, Kubernetes cluster, or HPC login node over SSH.$ hanzo science attach --ssh lab_cluster --program prog_ai_kai_latency_v4

Academic, National Lab & Open Science Grants

Discounted Team seats for universities, research hospitals, and nonprofit institutes, plus up to $30,000 in model credits and $2,000 in on-demand GPU compute for 50 open research programs.

Agentic R&D Operations on hanzo.team

Coordinate @Sci, @Dev, @Data, @QA, @Cyber, @Doc, and @Mana in shared channels with live notebooks, targets + crew Slurm DAGs, RO-Crate provenance, and 7-dimension Reviewer sign-off.hanzo.team · #research · Enso + Zen + Kai + 7/7 Verified ✓