Wave 5 — approachability & front door: CLI — discover --preset (physics/ecology/epidemiology/finance/general) + presets catalog; explicit flags override presets. library becomes a provenance-aware workspace: world/data content hashes, columns, config, tags captured on add; library search + compare over the workspace (backward-compatible index). SDK — lawsynth.recipes (parity with CLI presets) + Study.discover(recipe=..., **overrides); three runnable domain demos. Van der Pol shows the payoff: general recipe misses the cubic term, mechanics recipe recovers -x+v-x^2*v (R2 0.99995). Docs — a real product README (hero + Lorenz GIF, CLI+Python quickstart, feature matrix by product loop, surfaces + HTTP endpoints, honest tone); refreshed getting-started pages and a CHANGELOG Unreleased section. All flags verified against the shipped CLI usage. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
3.1 KiB
Quickstart: the core loop end to end
This walks the full discover → understand → use → share loop on a two-state
system. Substitute your own columns for x,y.
1. Prepare observations
Create a numeric CSV with a header, a strictly increasing finite time column, and
one or more finite state columns. For example observations.csv:
time,x,y
0,1,0
0.05,0.998,0.099
0.10,0.990,0.197
CSV, TSV, and a numeric Parquet subset are accepted (.csv, .tsv, .parquet).
2. Discover the law system
lawsynth discover observations.csv \
--time time --state x,y --output world.lsworld
discover writes a portable .lsworld bundle and prints the fit (mean squared
error) and complexity. Useful options: --degree N, --threshold VALUE,
--solver stlsq|sr3, --trigonometric, --rational, one derivative estimator
(--spline, --spectral, --savgol-window ODD_N, or --tvreg-lambda VALUE),
--smooth-radius N, --bootstrap REPLICATES, --symbolic-depth N, and the
optional stages --regimes, --pareto, --refine, --causal.
3. Understand what it found
lawsynth explain world.lsworld
lawsynth inspect world.lsworld
explain prints the readable laws, their dominant terms, the dependency structure,
and the assumptions a result is contingent on. inspect reports state, variable,
and parameter counts and distinguishes continuous from discrete worlds.
4. Use it: simulate and forecast
# Simulate over a fixed window
lawsynth simulate world.lsworld --initial x=1 --initial y=0 \
--start 0 --end 10 --step 0.05
# Forecast beyond the observed window
lawsynth forecast world.lsworld --horizon 20 --step 0.05 --output forecast.csv
Interventions are first-class and scheduled: --parameter-at TIME:NAME=VALUE and
--input-at TIME:NAME=VALUE on simulate, and --intervene NAME=VALUE@TIME on
forecast. Each targets a parameter or a non-state input at a given time — for
example --intervene beta=0.2@5 on an SIR world.
5. Compare
Diff two worlds directly:
lawsynth compare world.lsworld other.lsworld --html compare.html
Or overlay named what-if scenarios (each a set of scheduled interventions) against a baseline over one horizon:
lawsynth scenarios world.lsworld --horizon 20 --step 0.05 \
--scenario mitigated:beta=0.2@5 --scenario surge:beta=0.6@5 --html scenarios.html
The Python Study API also offers scenario boards keyed by initial-condition
overrides (study.add_scenario("hot", interventions={"x": 2.0})) — see
python.md.
6. Share
lawsynth report world.lsworld --output report.html
lawsynth export world.lsworld --format python --output world_model.py
report is a single self-contained HTML file — rendered equations, fit, regime
timeline, uncertainty bands, and inline SVG charts, with no server or external
assets. export emits a dependency-free Python module, a LaTeX align* block, or a
documented JSON description.
Discovery is deterministic for identical input and options: rerun any step and you get exactly the same world, report, and forecast.