nextstat 0.9.7

High-performance statistical inference engine — MLE, Bayesian, survival, GLM, time series. Rust core with Python bindings.

NextStat

NextStat is a high-performance statistical fitting toolkit for High Energy Physics (HEP), implemented in Rust with Python bindings.

The God Run (toy-based CLs)

• Model: S+B HistFactory (synthetic), 50 channels × 4 bins, 201 parameters (mu + 200 nuisances)
• Task: CLs via toy-based q~_mu
• Load: 10,000 toys (b-only) + 10,000 toys (s+b)
• Machine: Apple M5 (arm64), macOS-26.2-arm64-arm-64bit-Mach-O
• Versions: nextstat 0.1.0, pyhf 0.7.6, Python 3.13.11
• Recorded: 2026-02-07 (UTC)
• Commit: 88d57856

Tool	Wall time	Speedup
NextStat (Rayon)	3.47 s	1.0×
pyhf (multiprocessing, 10 procs)	50m 11.7s	868.0×

Reproduce:

PYTHONPATH=bindings/ns-py/python ./.venv/bin/python scripts/god_run_benchmark.py --n-toys 10000

What You Get

• pyhf JSON compatibility (HistFactory-style workspaces)
• Native HS3 (HEP Statistics Serialization Standard) v0.2 support — load ROOT 6.37+ HS3 JSON directly, auto-detected alongside pyhf
• Native ROOT TTree reader with mmap I/O, rayon-parallel basket decompression, and columnar extraction — no ROOT C++ dependency
• Ntuple-to-workspace pipeline: ROOT ntuples → histograms → HistFactory workspace (TRExFitter replacement)
• Expression engine for string-based selections and weights ("njet >= 4 && pt > 25.0")
• Negative log-likelihood (Poisson + constraints), including Barlow-Beeston auxiliary terms
• Maximum Likelihood Estimation (L-BFGS-B) with uncertainties via (damped) Hessian-based covariance + diagonal fallback
• NUTS sampling surface (generic Posterior API) + optional ArviZ integration
• SIMD kernels, Rayon parallelism, Apple Accelerate (vDSP/vForce), and optional GPU acceleration (CUDA for NVIDIA, Metal for Apple Silicon)
• Rust library, Python package (PyO3/maturin), and a CLI
• Implemented packs: regression/GLM, hierarchical models, time series (Kalman/EM/forecast), econometrics/causal helpers, and PK/NLME baselines

Docs

• Docs index: docs/README.md
• Tutorials (end-to-end): docs/tutorials/README.md
• References (CLI/Python/Rust/Server/Tools): docs/references/
• Demo: Physics Assistant (ROOT -> anomaly scan -> p-values + plots): docs/demos/physics-assistant.md

Quickstart

Install (Rust)

cargo add ns-core ns-inference ns-compute

Install (Python)

pip install nextstat

Build From Source

git clone https://github.com/NextStat/nextstat.io.git
cd nextstat.io

# Rust workspace
cargo build --release

# Python bindings (editable dev install)
cd bindings/ns-py
maturin develop --release

Try the Playground (WASM)

Run asymptotic CLs upper limits (Brazil bands) in the browser (no Python, no server) using a pyhf-style workspace.json.

From the repo root:

rustup target add wasm32-unknown-unknown
cargo install wasm-bindgen-cli --version 0.2.108

make playground-build-wasm
make playground-serve

Open http://localhost:8000/ and drag & drop a workspace.json (example: playground/examples/simple_workspace.json).

Usage

Rust API

use ns_inference::mle::MaximumLikelihoodEstimator;
use ns_translate::pyhf::{HistFactoryModel, HistoSysInterpCode, NormSysInterpCode, Workspace};

let json = std::fs::read_to_string("workspace.json")?;
let workspace: Workspace = serde_json::from_str(&json)?;

// Default interpolation (NextStat "smooth" defaults): NormSys=Code4, HistoSys=Code4p.
// For strict HistFactory/pyhf defaults, use Code1/Code0:
let model = HistFactoryModel::from_workspace_with_settings(
    &workspace,
    NormSysInterpCode::Code1,
    HistoSysInterpCode::Code0,
)?;

let mle = MaximumLikelihoodEstimator::new();
let result = mle.fit(&model)?;

println!("Best-fit params: {:?}", result.parameters);
println!("NLL at minimum: {}", result.nll);

Python API

import json

import nextstat

workspace = json.loads(open("workspace.json").read())
model = nextstat.from_pyhf(json.dumps(workspace))
result = nextstat.fit(model)

poi_idx = model.poi_index()
print("POI index:", poi_idx)
print("Best-fit POI:", result.bestfit[poi_idx])
print("Uncertainty:", result.uncertainties[poi_idx])

Population PK (nlme_foce) with multi-cpt FO/ITS/IMP

import nextstat

times = [0.5, 1.0, 2.0, 4.0, 8.0] * 4
subject_idx = [0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3]
y = [8.0, 6.2, 4.1, 2.6, 1.2] * 4

fit_fo = nextstat.nlme_foce(
    times,
    y,
    subject_idx,
    4,
    model="2cpt_iv",
    method="fo",
    doses=[120.0],  # length-1 broadcasts to all subjects
    theta_init=[1.2, 15.0, 0.8, 20.0],
    omega_init=[0.2, 0.2, 0.2, 0.2],
    error_model="additive",
    sigma=0.1,
)

fit_imp = nextstat.nlme_foce(
    times,
    y,
    subject_idx,
    4,
    model="3cpt_iv",
    method="imp",
    doses=[120.0],
    theta_init=[1.1, 14.0, 0.7, 18.0, 0.5, 28.0],
    omega_init=[0.2] * 6,
    imp_n_iter=5,
    imp_n_samples=100,
    error_model="additive",
    sigma=0.1,
)

Unbinned (event-level) API

Compile an event-level likelihood from an unbinned_spec_v0 JSON/YAML file and run fits/scans:

import nextstat

analysis = nextstat.unbinned.from_config("unbinned.json")
fit = analysis.fit()
print("NLL:", fit.nll)

scan = analysis.scan([0.0, 0.5, 1.0, 2.0])
print("mu_hat:", scan["mu_hat"])

cls = analysis.hypotest_toys(1.0, n_toys=2000, seed=42)
print("CLs:", cls)

Bayesian (NUTS) + ArviZ

Install optional deps:

pip install "nextstat[bayes]"

Run sampling and get an ArviZ InferenceData:

import json
from pathlib import Path

import nextstat

workspace = json.loads(Path("workspace.json").read_text())
model = nextstat.from_pyhf(json.dumps(workspace))

idata = nextstat.bayes.sample(
    model,
    n_chains=2,
    n_warmup=500,
    n_samples=1000,
    seed=42,
    target_accept=0.8,
)

print(idata)

Viz (CLs Brazil bands, profile scans)

Install optional deps:

pip install "nextstat[viz]"

Compute artifacts and plot (matplotlib):

import json
import numpy as np
from pathlib import Path

import nextstat

workspace = json.loads(Path("workspace.json").read_text())
model = nextstat.from_pyhf(json.dumps(workspace))

scan = np.linspace(0.0, 5.0, 101)
cls_art = nextstat.viz.cls_curve(model, scan, alpha=0.05)
nextstat.viz.plot_cls_curve(cls_art, title="CLs Brazil band")

mu = [0.0, 0.5, 1.0, 2.0]
prof_art = nextstat.viz.profile_curve(model, mu)
nextstat.viz.plot_profile_curve(prof_art, title="Profile likelihood scan")

Ntuple → Workspace (TRExFitter replacement)

use ns_translate::NtupleWorkspaceBuilder;

let ws = NtupleWorkspaceBuilder::new()
    .ntuple_path("ntuples/")
    .tree_name("events")
    .measurement("meas", "mu")
    .add_channel("SR", |ch| {
        ch.variable("mbb")
          .binning(&[0., 50., 100., 150., 200., 300.])
          .selection("njet >= 4 && pt > 25.0")
          .data_file("data.root")
          .add_sample("signal", |s| {
              s.file("ttH.root")
               .weight("weight_mc * weight_sf")
               .normfactor("mu")
          })
          .add_sample("background", |s| {
              s.file("ttbar.root")
               .weight("weight_mc * weight_sf")
               .normsys("bkg_norm", 0.9, 1.1)
               .weight_sys("jes", "weight_jes_up", "weight_jes_down")
               .tree_sys("jer", "jer_up.root", "jer_down.root")
               .staterror()
          })
    })
    .build()?;  // → Workspace (same type as pyhf JSON path)

No ROOT C++ dependency. ~8.5x faster than uproot+numpy on the full pipeline.

Low-level TTree access

use ns_root::RootFile;

let file = RootFile::open("data.root")?;
let tree = file.get_tree("events")?;

// Columnar access
let pt: Vec<f64> = file.branch_data(&tree, "pt")?;
let eta: Vec<f64> = file.branch_data(&tree, "eta")?;

// Expression engine
let expr = ns_root::CompiledExpr::compile("pt > 25.0 && abs(eta) < 2.5")?;

CLI

nextstat fit --input workspace.json
nextstat --interp-defaults pyhf fit --input workspace.json   # NormSys=Code1, HistoSys=Code0
nextstat hypotest --input workspace.json --mu 1.0 --expected-set
nextstat hypotest-toys --input workspace.json --mu 1.0 --n-toys 10000 --seed 42 --threads 0
nextstat hypotest-toys --input workspace.json --mu 1.0 --n-toys 10000 --gpu cuda   # NVIDIA GPU (f64)
nextstat hypotest-toys --input workspace.json --mu 1.0 --n-toys 10000 --gpu metal  # Apple Silicon GPU (f32)
nextstat upper-limit --input workspace.json --expected --scan-start 0 --scan-stop 5 --scan-points 201
nextstat version

Documentation

• Tutorial index: docs/tutorials/README.md
• Python API reference: docs/references/python-api.md
• Rust API reference: docs/references/rust-api.md
• CLI reference: docs/references/cli.md
• Playground (browser/WASM): docs/references/playground.md (and playground/README.md)

Architecture

NextStat follows a "clean architecture" style: inference depends on stable abstractions, not on specific execution backends.

┌─────────────────────────────────────────────────────────────────┐
│                        HIGH-LEVEL LOGIC                          │
│  ns-inference (MLE, Profile Likelihood, Hypothesis Tests, ...)   │
│  - depends on core types and model interfaces                     │
└─────────────────────────┬───────────────────────────────────────┘
                          │ depends on abstractions
┌─────────────────────────┴───────────────────────────────────────┐
│                      ns-core (interfaces)                        │
│  - error types, FitResult, traits                                 │
└─────────────────────────┬───────────────────────────────────────┘
                          │ implemented by
┌─────────────────────────┴───────────────────────────────────────┐
│                    LOW-LEVEL IMPLEMENTATIONS                     │
│  ns-translate (pyhf + ntuple → Workspace)  ns-compute (SIMD/CUDA/Metal) │
│  ns-ad (dual/tape AD)  ns-root (ROOT I/O, TTree, expressions)    │
└─────────────────────────────────────────────────────────────────┘

Project Layout

nextstat/
├── crates/
│   ├── ns-core/         # Core types, traits, error handling
│   ├── ns-compute/      # SIMD kernels, Apple Accelerate, CUDA/Metal batch NLL+grad
│   ├── ns-ad/           # Automatic differentiation (dual/tape)
│   ├── ns-root/         # Native ROOT file reader (TH1, TTree, expressions, filler)
│   ├── ns-translate/    # Format translators (pyhf, HS3, HistFactory XML, ntuple builder)
│   ├── ns-inference/    # MLE, NUTS, CLs, GLM, time series, PK/NLME
│   ├── ns-viz/          # Visualization artifacts
│   └── ns-cli/          # CLI binary
├── bindings/
│   └── ns-py/           # Python bindings (PyO3/maturin)
├── docs/
│   ├── legal/
│   ├── plans/
│   └── references/
└── tests/

Development

Requirements

• Rust 1.93+ (edition 2024)
• Python 3.11+ (for bindings)
• maturin (for Python bindings)

Build and Test

# Build
cargo build --workspace

# Build with CUDA support (requires nvcc)
cargo build --workspace --features cuda

# Build with Metal support (Apple Silicon, macOS)
cargo build --workspace --features metal

# Tests (default features)
cargo test --workspace

# Tests including optional backends (CUDA requires nvcc)
cargo test --workspace --all-features

# Opt-in slow Rust tests (toys, SBC, NUTS quality gates)
make rust-slow-tests

# Very slow (release) regression check
make rust-very-slow-tests

# Format and lint
cargo fmt --check
cargo clippy --workspace -- -D warnings

Python Tests

Local test runs should use the repo venv (it pins a Python version compatible with the built extension).

# Run fast Python tests (parity + API contracts)
PYTHONPATH=bindings/ns-py/python ./.venv/bin/python -m pytest -q -m "not slow" tests/python

# Run slow toy regression tests (opt-in)
PYTHONPATH=bindings/ns-py/python NS_RUN_SLOW=1 NS_TOYS=200 NS_SEED=0 ./.venv/bin/python -m pytest -q -m slow tests/python

Benchmarks

# Compile and run all benches
cargo bench --workspace

# Common entry points
cargo bench -p ns-translate --bench model_benchmark
cargo bench -p ns-translate --bench nll_benchmark
cargo bench -p ns-compute --bench simd_benchmark
cargo bench -p ns-inference --bench mle_benchmark
cargo bench -p ns-inference --bench hypotest_benchmark
cargo bench -p ns-ad --bench ad_benchmark
cargo bench -p ns-core --bench core_benchmark

Details (quick mode, baselines, CI workflows): docs/benchmarks.md.

Apex2 Baselines (pyhf + P6 GLM)

Record a reference baseline (writes JSON under tmp/baselines/ with a full environment fingerprint):

make apex2-baseline-record

Compare current HEAD vs the latest recorded baseline (writes tmp/baseline_compare_report.json):

make apex2-baseline-compare

Pre-release gate runbook: .internal/docs/internal/release-gates.md.

Documentation

• White paper (Markdown): docs/WHITEPAPER.md
• White paper (PDF): built by python3 scripts/build_whitepaper.py and attached to GitHub Releases on tags (v*)
• Internal plans/design notes are maintained outside this public repository.

Contributing

See CONTRIBUTING.md. All commits must include DCO sign-off (git commit -s).

License

NextStat uses a dual-licensing model:

• Open Source: LICENSE (AGPL-3.0-or-later)
• Commercial: LICENSE-COMMERCIAL

Contact

• Website: https://nextstat.io
• GitHub: https://github.com/NextStat/nextstat.io