scientific-computing-system-2.0

This is the NumPy build of scientific-computing-system, not a separate product: pick this one for NumPy/SciPy/pandas/matplotlib, the other for readable pure Python with zero runtime dependencies.

Name and version

2.0 names the NumPy generation of the project, not the release number. The current release is 5.4.5, and PyPI has never published a 2.x version.

PyPI and npm project names are fixed after publication, so the -2.0 suffix stays even though the version has moved past it.

53 importable names, 535 public exports, 1,927 tests, 100% branch coverage, MIT. Every count here is measured from the tree by scripts/make_promo.py; CI re-measures and rejects drift.

Reproducibility drift is a result, not hidden state

guided-fit saves the random seed, input hashes and exact Python/NumPy/SciPy/pandas/Matplotlib versions in its manifest. A later guided-fit-rerun compares the saved analysis with the new one instead of silently treating the rerun as equivalent.

When something moves, the warning is concrete. The CLI reports the exact saved and current dependency versions, and confidence-interval drift includes the relative change, the largest bound shift, and the saved and rerun interval bounds:

warning   rerun differs materially from the saved analysis
detail    runtime version changed: scipy <saved-version> -> <current-version>
detail    confidence intervals changed materially for experiment: relative=<...>; max_bound_shift=<...>; saved=[[...]]; rerun=[[...]]

The placeholders above are documentation only; an actual rerun prints the measured versions and interval values from that run. Version drift is surfaced even when the fit still agrees numerically, while confidence-interval drift is checked independently from RMSE and parameter drift.

Real archived replay (2026-10-06). A saved fit from Python 3.12.14 / NumPy 1.26.4 / SciPy 1.11.4 was replayed with the same input hash and seed under Python 3.12.14 / NumPy 2.3.5 / SciPy 1.18.1. The real rerun emitted both exact version-change warnings. The largest 95% CI-bound shift was 1.1747583669e-4, relative CI drift was 0.003364917%, and the reliability label stayed reliable -> reliable; therefore the environment changed but the numerical result did not cross the 5% material-drift threshold. See the case study and the archived result JSON.

Installation

Python 3.10+. PyPI is the install path:

pip install scientific-computing-system-2.0

From a clone: pip install -e ".[dev,docs]" (pytest plus the MkDocs toolchain).

Optional C kernels. _fast_kmeans and _fast_pagerank are compiled when a toolchain is available and fall back silently to pure NumPy otherwise, so CDS_PURE=1 pip install . is fully functional but slower on those two operations.

npm (optional). A same-named npm package is a thin Node launcher shim — five files, ~8 KB, no Python — that calls python -m cds2.cli. You still need the PyPI install, so most people can skip it: docs/npm.md.

Supply chain: PyPI and npm releases are attested. Every file of every PyPI release from 5.1.0 through 5.4.5 carries a PEP 740 provenance attestation: the per-file pypi.org/integrity/.../<file>/provenance endpoint returns 200 (all 330 files checked on 2026-10-22). The npm launcher caught up on 2026-10-23: 5.4.5 is the first npm version with a Sigstore attestation (200, SLSA provenance v1); every earlier npm version, including 5.4.4, returns 404 and has no attestations field. Pin digests where provenance is absent. What the project does guarantee (reproducible wheel, an sdist that really ships the native kernels, version lockstep, cross-platform smoke) and what a consumer can check: Supply chain.

Quick start

Executed against scientific-computing-system-2.0==5.2.5:

import cds2

print(cds2.linalg.solve([[3.0, 1.0], [1.0, 2.0]], [9.0, 8.0]))  # [2. 3.]
print(cds2.montecarlo.pi_estimate(n=100_000, seed=42))  # 3.13776
print(cds2.infotheory.entropy([0.25, 0.25, 0.25, 0.25]))  # 2.0
print(cds2.graph.pagerank(cds2.graph.from_edges(4, [(0, 1), (0, 2), (1, 3), (2, 3)])))
# [0.1375043  0.19594362 0.19594362 0.47060846]

Guided fitting: the differentiating feature

cds2.guided_fit turns a CSV into a reproducible, inspectable model fit. Unlike a bare curve_fit, the model is recommended and then confirmed by you, and the run is written to a replayable manifest.

cds2 guided-fit data.csv --x time --y response \
    --model exponential --missing drop --outliers exclude --report pdf
cds2 guided-fit-rerun guided-fit-results/guided_fit_manifest.json

Five model families (linear, quadratic, exponential, power, logistic), explicit missing-data (ask/drop/interpolate) and outlier (ask/keep/exclude) policies, PDF/HTML/Markdown reports, and a manifest carrying the input hash so a rerun proves it used the same data. See case studies.

Command line

All ten subcommands, from cds2 --help:

cds2 info                      show version information
cds2 stats 1,2,3,4,5           descriptive statistics
cds2 integrate sin --a 0 --b 3.14159
cds2 linsolve --a "3,1;1,2" --b "9,8"
cds2 entropy "0.25,0.25,0.25,0.25"
cds2 units 5 --from-unit km --to-unit mile
cds2 solve --coeffs "1,-5,6"
cds2 plot 1,3,2,5,4 --file out.png
cds2 guided-fit data.csv --x time --y response
cds2 guided-fit-rerun guided-fit-results/guided_fit_manifest.json

python -m cds2 works and is equivalent to the cds2 console script, so either invocation is fine:

cds2 info
python -m cds2 info

Modules

47 flat modules plus 6 subpackages; 47 generated API pages. * marks deprecated modules.

Examples

21 runnable scripts in examples/, all executed by the examples-smoke CI job on every push. Try bayesian_inference.py (Metropolis-Hastings vs an analytic conjugate result), experiment_fitting.py (Michaelis-Menten fit, permutation test on residuals) and signal_denoising.py (Butterworth filter, Welch spectra). Catalogue: Examples · Recipes.

Performance, measured honestly

benchmarks/results.json and docs/benchmarks.md are a real run, but a stale one: cds2 3.0.0, commit 6ea5a02, 2026-08-22, while the release is 5.4.5. Read them as "3.0.0 versus the baseline libraries on one machine", not as a claim about today.

Of the 13 measured cases, CDS v2 was slower in 4: dataframe summary 1.82x, solve 800x800 1.22x, solve 8x8 x300 1.17x and welch 1.03x. The pandas row returns strictly more information (per-column nulls and uniques), so part of that premium is extra work. The clearest win is PageRank at 0.15x versus NetworkX. There is no blanket speedup claim, and benchmark-methodology.md sets out the limits.

Documentation map

Contributing

pip install -e ".[dev,docs]"
pytest --cov=cds2 --cov-fail-under=100   # the gate is fail_under=100
mkdocs serve                              # docs at http://127.0.0.1:8000

Coverage below 100% fails CI, so new code needs tests on success and error paths. Full rules in CONTRIBUTING.md.

Deprecated modules

cds2.special and cds2.distributions are deprecated since 4.3.0. Because __init__.py imports both eagerly, a bare import cds2 emits both DeprecationWarnings even if you never touch them:

DeprecationWarning: cds2.special is deprecated since 4.3.0 ...
DeprecationWarning: cds2.distributions is deprecated since 4.3.0 ...

They still work; prefer scipy.special and scipy.stats directly. Migration notes: Deprecated modules.

Project files

CHANGELOG.md · CONTRIBUTING.md · SECURITY.md · CODE_OF_CONDUCT.md · CITATION.cff · codemeta.json · LICENSE (MIT)

Related projects

Part of the Furox scientific-computing portfolio: independently versioned projects that pair agent-facing interfaces (MCP servers and Agent Skills) with deterministic, fail-closed verification.

• scientific-computing-system — zero-dependency Python numerics with run manifests.
• scientific-toolkit — MCP gateway over nine research tools.
• axiomize — unit-checked modeling engine with SBML, CellML and Modelica export.
• axiomize-quantum-skills-2.0 — modeling engine plus quantum-inspired reasoning bundle.
• quantum-reasoning-skill — agent skill that keeps competing hypotheses alive.
• plan-auditor — evidence-based completion gate for coding agents.
• nomosguard — deterministic security decision layer with a hash-chained ledger.
• eq-layer — interaction-policy layer that decides how an agent speaks.
• literature-evidence-mcp — literature MCP server with a hash-chained evidence ledger.
• repocraft — byte-deterministic project scaffolding.

License

MIT.