Adds docs/getting-started/faq.md targeting real search queries people ask when evaluating equation-discovery / symbolic-regression tools (what SINDy is, how LawSynth compares to PySINDy/PySR, whether it's free/offline, what a .lsworld bundle is, and honest current limitations) — claim-gated against the actual CLI, README, and docs/research/limitations.md. Also fixes two unresolved git merge-conflict markers left in apps/docs-site/package.json and src/site.ts from a prior merge (ef2abe9), which left package.json as invalid JSON and broke `npm run build/test/render` for the lawsynth.dev static-site pipeline. Resolved both in favor of the feat/release-cut-2026-09-06 side (keeps seo_titles.test.js in the test script and fallbackDescription() in renderPage, which seo_titles.ts already exports and the rest of the file already relies on). Build, full test suite (15/15), and a deterministic re-render all pass after the fix. Rendered and deployed via `npx wrangler deploy`; new page verified live at https://lawsynth.dev/docs/getting-started/faq (200, correct title/meta, listed in sitemap.xml), and all referenced internal doc links verified 200. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
6.0 KiB
Frequently asked questions
Short, direct answers to the questions people ask before trying LawSynth — what equation discovery is, how it compares to SINDy and symbolic-regression tools, and what it can and cannot do today.
What is "equation discovery from time-series data"?
It is the process of recovering an explicit mathematical model — usually a
system of differential equations dx/dt = f(x) — directly from measured
data, instead of hand-deriving it from theory or fitting an opaque black-box
model. You supply a time column and one or more numeric state columns; a
discovery algorithm searches a library of candidate terms (polynomial,
trigonometric, rational, …) and returns the sparsest combination that fits
the observed dynamics. The output is a readable equation, not a set of
weights — see core concepts and the
sparse regression guide.
What is SINDy, and is LawSynth "like SINDy"?
SINDy (Sparse Identification of Nonlinear Dynamics) is the leading published
method for this problem: it builds a library of candidate terms and solves a
sparse regression (originally sequentially-thresholded least squares) to keep
only the few terms that explain the measured derivatives. LawSynth implements
the same family of sparse-regression methods — STLSQ and SR3 solvers, with
stability selection — as one of its discovery solvers (lawsynth discover --solver stlsq|sr3), alongside an optional bounded symbolic-search stage.
In plain terms: SINDy is a method, PySINDy is the reference Python library
for it, and LawSynth is a separate, Rust-core, deterministic implementation
of the same underlying idea, delivered as a CLI/SDK/Studio product rather
than a library you import into a notebook. See the
sparse regression methods and the
full competitive comparison for a
line-by-line feature comparison against PySINDy, PySR, Operon, SciML,
gplearn, and AI-Feynman.
Is there a free, open-source alternative to PySINDy or symbolic-regression tools?
Yes. LawSynth is open source under Apache-2.0 and runs entirely on your own
machine. The core engine is Rust with a Python SDK (maturin/pyo3) and a
TypeScript Studio UI. There is no account, no API key, and no network call
during discovery — cargo is configured offline and the CLI never sends
your data anywhere. See installation
and the determinism guide for what "offline and
deterministic" means in practice.
How is LawSynth different from PySR, Operon, or other symbolic-regression libraries?
Pure symbolic-regression tools (PySR, Operon, gplearn) search for a single
best-fit formula and are typically stochastic — rerunning the same search
can return a different expression. LawSynth's discovery is deterministic:
identical inputs and options reproduce the identical .lsworld bundle, bit
for bit, every time (lawsynth discover writes the bundle; lawsynth inspect reads it back). LawSynth also targets dynamics specifically
(systems of dx/dt = f(x) you can simulate forward), returns a Pareto front
of candidates ranked by fit versus complexity instead of one opaque score,
and ships the full loop — discover, explain, simulate, intervene, compare,
and share a signed, portable bundle — rather than only a formula string. The
competitive analysis documents where
each tool still leads (for example, PySINDy's weak/integral-form solvers and
PySR's distributed search are capabilities LawSynth does not yet have).
Can I discover a governing equation from a CSV without writing code?
Yes, from the command line: lawsynth discover observations.csv --time t --state x,y --output model.lsworld reads a CSV with a header and a strictly
increasing numeric time column, and writes a portable .lsworld bundle
reporting the leading candidate's mean-squared error and complexity. No
Python, Rust, or notebook code is required for this step. CSV, TSV, and a
constrained numeric Parquet subset are accepted. See the
CSV-to-model cookbook and the
discover CLI reference for every option.
What is a .lsworld bundle?
It is LawSynth's portable output format: a single file holding the
discovered law system, its parameters, provenance, and (optionally) a
self-contained HTML report. lawsynth inspect reads it back, lawsynth simulate runs it forward, and lawsynth intervene applies scheduled
interventions to it — so the same artifact you get from discovery is what
you share, version, and re-run later. See
Worlds and
organize & share your work.
What does LawSynth cost?
Nothing. It is free and open source (Apache-2.0); there is no paid tier, account, or usage limit — you build and run the CLI/SDK locally from the GitHub repository.
What can't LawSynth do (yet)?
Stated plainly, so you can decide if it fits before trying it: discovery does not infer missing values, latent states, measurement-error models, or automatic causal identification — preprocessing and scientific judgment remain your responsibility. The native Parquet reader only accepts an uncompressed, flat, PLAIN-encoded numeric subset (it rejects other encodings outright rather than risk silently misreading them). A computed simulation trajectory is a numerical result, not a stability proof or uncertainty certificate. LawSynth also does not yet publish an SRBench benchmark result or ship an sklearn-compatible estimator — both are open, tracked gaps, not finished features. Full detail: limitations and failure cases.
Where do I start?
Follow the quickstart for the full discover → understand → use → share loop on a two-state example, or jump straight to installation if you already have a CSV in hand.