Skip to content

About

Scientific modeling engine (versioned Model IR, dimensional validation, uncertainty) plus quantum-inspired multi-branch reasoning. Python, MCP and REST. MIT.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

210 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Axiomize Quantum Skills — 2.0

This is not a separate product. It ships axiomize and quantum-reasoning-skill together. Install those repositories when you need only one of them. This repository is not a second MCP registration.

Current package line: 1.2.6

License Python CI PyPI npm Docs

Download numbers live on the package pages themselves, where they are always current: PyPI · npm

The npm 1.2.4 tarball carries Sigstore attestations, including a full SLSA v1 build provenance statement. The PyPI 1.2.4 upload was made manually and has no PEP 740 attestation — do not describe that release as attested. From 1.2.5 onward, artifacts on both registries are published by the attested release workflow (OIDC trusted publishing). The npm package is a thin Node launcher shim, not the engine: six files, about 66.7 kB unpacked (about 25 kB as a tarball), one job — spawn the Python CLI. So install the Python package first. Details, the three mechanisms they are easy to confuse with, and a verification recipe are in docs/publishing-checklist.md.

Documentation: https://furox-art.github.io/axiomize-quantum-skills-2.0/

A scientific modeling engine with explicit units, dimensional and numerical validation, calibration and uncertainty analysis — plus a reasoning protocol that refuses to commit to the first plausible answer.

Piece What it is Where
Axiomize Deterministic engine: versioned Model IR; algebraic, ODE, stochastic, PDE, optimization, control, network, Bayesian and causal execution; dimensional checks; fit; sensitivity; uncertainty; portable export src/axiomize/, CLI axiomize
Quantum Reasoning A protocol and deterministic branch controller that keeps several hypotheses alive, scores them against evidence, prunes weak ones, and collapses only when the evidence supports it src/axiomize/reasoning/, CLI axiomize-reason

"Quantum" here is a metaphor for deferring commitment, not quantum computation. It runs on classical models and classical hardware.

See it work

The reference SIR model, animated. beta = 0.3, gamma = 0.1, so R0 = 3.0.

Animated SIR epidemic curve, R0=3.0

The plotted curve comes from the axiomize-validate --plot flag shown in step 3. The static snapshot at full resolution, and the LaTeX sample it converts to, are in docs/index.md.

Quick start (verified against 1.2.0)

The current package line is 1.2.5. The commands below were last re-verified against 1.2.0 at commit 9c2990c; that verification is a historical record, not a claim about 1.2.5.

Install

pip install axiomize-quantum-skills-2.0. Python 3.10+. NumPy, SciPy, SymPy, NetworkX, statsmodels, Matplotlib, Control, CVXPY, CasADi and z3-solver come along automatically.

1. See what is actually installed

axiomize tools reports only backends that genuinely imported:

sympy        1.14.0     available
scipy        1.18.1     available
statsmodels  0.15.0     available
cvxpy        1.9.3      available
jax          unknown    UNAVAILABLE (No module named 'jax')
lean         Lean (version 4.30.0, x86_64-w64-windows-gnu, ...) available
pymc                    UNAVAILABLE (pymc not installed)
fenics                  UNAVAILABLE (fenics not installed)

The full list has 13 entries. axiomize capabilities is the machine-readable equivalent and leads with "axiomize_version": "1.2.6".

2. Run the install-safe benchmark suite

axiomize benchmark → {"status": "PASS", "passed": 12, "total": 12}.

3. Validate a model, and plot it

axiomize-validate --model sir --beta 0.3 --gamma 0.1 --N 1000000 --plot sir.png
=== SIR validation ===
horizon                = 180 days  (final-size theory is the t->infinity limit)
R0                     = 3.000  (outbreak)
Peak infected          = 300,465 at day 61.4
Final size (simulated) = 0.9404
Final size (theory)    = 0.9405
Theory match           = True

--- sanity checks ---
population_conserved                PASS
compartments_nonnegative            PASS
R_monotonic_increase                PASS

plot saved -> sir.png

axiomize-validate also takes --model gillespie (extinction risk) and --model queue (Erlang-C staffing cliff), plus --sweep and --seed. The --plot flag is what writes a PNG; axiomize validate --N 1000000 emits JSON only and writes no file. That JSON form cross-checks two independent methods (cross_validation PASS, agreeing to 1e-16) then declares numeric_theory_check INCONCLUSIVE, because the finite horizon has not reached the asymptotic SIR state — see docs/tutorial.md for it in full.

4. Work with your own model

axiomize model --input-json docs/quickstart-model-ir.json --action validate

docs/quickstart-model-ir.json is a complete, copy-pasteable request. To start from a plain-language idea instead, axiomize intake "<your idea>" asks clarifying questions and recommends a depth. Note axiomize model requires --input-json and has no --idea flag.

5. Score a reasoning branch

axiomize-reason score --evidence 0.9 --verification 0.85 → {"branch": "branch-1", "can_collapse": false, "reason": "leader score below collapse threshold"} — it declines to collapse, which is correct: a branch needs verification and margin, not just high evidence. The other two subcommands are select (rank model candidates from a {"candidates": [...]} file) and width (an uncertainty band from a list of scores).

Every command above was run against axiomize-quantum-skills-2.0 1.2.0 at commit 9c2990c on CPython 3.12.10 (Windows); that is a historical verification record. The current package line is 1.2.5. docs/tutorial.md is the full walkthrough with real output for each step.

Interfaces

Interface Entry point
Main CLI axiomize — 14 subcommands: intake, policy, model, clean-data, compare-runs, solve, fit, validate, tools, capabilities, reproduce, benchmark, serve, mcp
Validation and data tools axiomize-validate (reference model sanity checks, --sweep, --plot), axiomize-fit (calibrate to a CSV, --selftest, --compare, --plot), axiomize-csv-check (data-quality pre-check before fitting)
Benchmarks and reports axiomize-benchmark (grade a report against a case), axiomize-to-latex (report → LaTeX/PDF), axiomize-index-reports (build a report index)
Parallel execution axiomize-sweep (--job sweep or --job mc) — the skill's subagent-dispatch pattern applied to parameter grids
Reasoning axiomize-reason — score, select, width
MCP (stdio) axiomize mcp — 34 tools; host config {"command": "axiomize", "args": ["mcp"]}
REST (v1) axiomize serve — loopback-only unless --allow-remote with a bearer token

All of them call the same core services, so validation behaviour never depends on the caller. docs/integrations.md has every subcommand, MCP tool and REST route.

The agent skills are plain Markdown protocols and install without the Python package:

git clone https://github.com/Furox-Art/axiomize-quantum-skills-2.0
cp -r axiomize-quantum-skills-2.0/skills/axiomize          ~/.config/opencode/skills/
cp -r axiomize-quantum-skills-2.0/skills/quantum-reasoning ~/.config/opencode/skills/

What you get

  • 15 mathematical perspectives under skills/axiomize/perspectives/, plus archetype and first-principles routing, and 5 report templates in skills/axiomize/templates/ — the 10-section report, parameters, assumptions, glossary, subagent brief.
  • 12 domain packs in packs/: curated bundles of which archetypes, lenses and examples matter in a given domain — epidemiology, operations, economics, ecology, project management, physics, chemistry, biology, climate, engineering, finance, probability, plus the index.
  • 18 worked examples in examples/, indexed by docs/example-gallery.md.
  • A 25-case closed-form benchmark corpus with stored reports, graded in CI, plus a 6-case reasoning corpus with 4 JSON schemas under src/axiomize/reasoning/benchmark/ (evaluate.py, validate_submission.py).
  • Versioned portable runs: axiomize.portable-bundle.v1 exports and run bundles carry a canonical Model IR SHA-256, so a run can be zipped, moved and re-inspected elsewhere. See docs/portable-export.md.
  • Optional backends, reported honestly by axiomize tools: a real Lean 4 proof checker (formal/lean_adapter.py) and an OpenAI-compatible provider layer (providers/) for model-backed steps, both behind the consumption guard in axiomize policy.
  • Explicit honesty rules: rejected lenses must carry a reason, every quantity carries a unit, and every model states what observation would falsify it.

Benchmarks

The 25-case corpus is graded in CI by skills/axiomize/tools/benchmark_runner.py (also axiomize-benchmark). Re-graded at commit 9c2990c (historical record), axiomize 1.2.0, CPython 3.12.10 (Windows), numpy 2.5.3 / scipy 1.18.1:

python skills/axiomize/tools/benchmark_runner.py --case physics-rc-discharge --report benchmarks/reports/physics-rc-discharge.md
automated score: 9/9 -> 10.0/10 (PASS; human rubric layer still required)

All 25 cases pass at 10.0/10, each against its own check count (8 to 11). This is a contract-compliance measurement, not a measure of modeling quality: the automated layer checks structure — literal alternative lists present, lens count, unit column, violation-consequence column, falsifiability section, numeric oracle within tolerance. A structurally perfect report about a wrong model still scores 10.0/10. The human rubric layer in benchmarks/rubric.md is not recorded for these cases; see docs/benchmark-results.md for the provenance-stamped run.

Honest limits

  • The engine does not choose a model for you. axiomize model --action plan ranks candidate families and then returns NEEDS_MODEL_IR; deterministic execution starts only after you supply an explicit Model IR.
  • Optional backends (pymc, jax, FEniCS/DOLFINx) are reported UNAVAILABLE unless actually installed. The [full] extra installs pymc and jax only — FEniCS/DOLFINx has no pip extra and must be installed separately. [playground] adds gradio and pandas.
  • Arbitrary code and theorem elaboration are not an OS sandbox. They require explicit trust. See SECURITY.md and docs/security.md.
  • The npm package is a launcher shim, not a second implementation. It is 6 files, about 66.7 kB unpacked (about 25 kB as a tarball), and it does one thing: spawn the Python CLI. If the Python package is not installed, the launcher propagates Python's ModuleNotFoundError, which exits 1; npx axiomize-quantum exits 127 only when no Python interpreter can be launched at all, and says which interpreter it tried. Set AXIOMIZE_PYTHON to pick one explicitly. There is no JavaScript engine here; use the axiomize console script or python -m axiomize.cli directly if you do not want the extra hop.
  • npm also still holds 2.0.0, published before the launcher fix, whose index.js is the old syntactically invalid one. It is not the latest dist-tag any more, so a plain npm install axiomize-quantum-skills-2.0 resolves to the current release. If you have pinned or installed 2.0.0, upgrade.

Documentation

The published site carries 20 pages; the mkdocs.yml nav is the source of truth.

Group Pages
Home index
Workflow workflow · adaptive · rigor · archetypes
Guides tutorial · worked examples · example gallery · integrations · portable export
Reference benchmark results
Security overview · versioning · CI contract · release checklist · maintenance · non-goals · 1.11.2 audit
Publishing publishing checklist

The workflow, rigor and archetypes pages are staged into docs/ from skills/axiomize/ at build time; their repository paths are skills/axiomize/{SKILL,adaptive-workflow,rigor,archetypes}.md.

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.

Contributing, citing, licence

CONTRIBUTING.md has the perspective and example contracts; CODE_OF_CONDUCT.md governs participation; SECURITY.md takes vulnerability reports. Cite via CITATION.cff; release history is in CHANGELOG.md. MIT — LICENSE, Copyright (c) 2026 Furkan.

About

Scientific modeling engine (versioned Model IR, dimensional validation, uncertainty) plus quantum-inspired multi-branch reasoning. Python, MCP and REST. MIT.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages