Relocate the two large generated planning docs (LawSynth_3161_Repository_Manifest.md 656K, LawSynth_Production_Architecture.md 118K) from the repo root into docs/internal/ so a public v0.1 checkout isn't ~770K of scaffolding at the top level. Update the three references in PRODUCT.md and docs/roadmap/ to the new paths and add docs/internal/README.md marking them as non-release artifacts. Note: untracked local datasets (creditcard.csv 144M, Financials.csv, heart.csv) were also deleted from the working tree; they are gitignored (*.csv) and were never tracked, so there is no git-history change here. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
5.5 KiB
LawSynth — Product
LawSynth turns time-series observations into executable mathematical worlds:
interpretable law systems you can read, simulate, stress-test, and share. This
document is the product view — what a user actually does with LawSynth — layered
on top of the engine and repository blueprint in ARCHITECTURE.md and
docs/internal/LawSynth_Production_Architecture.md.
Who it's for
- Scientists & engineers who have measurements and want a mechanistic model (equations, not a black box) they can trust and reason about.
- Analysts & quants who need to forecast, run interventions, and compare scenarios on a model they can inspect.
- Teams who want reproducible, shareable, local-first results — no data leaves the machine.
The core loop
observe (CSV) → discover (laws) → understand (explain) → use (simulate /
forecast / intervene) → compare → share (report / .lsworld bundle)
Every step is deterministic and offline. A discovery is a portable .lsworld
bundle; everything downstream operates on that artifact.
Product surfaces
| Surface | For | Entry point |
|---|---|---|
| CLI | power users, automation, CI | lawsynth <command> |
| Python SDK | notebooks, pipelines | import lawsynth |
| Studio | local interactive exploration | local browser interface (apps/studio) |
| Services | teams, self-hosting | API + gateway |
| Applied labs | bounded domain workflows | GridSynth and the native Rust information-diffusion app in this workspace |
Feature areas (product depth)
1. Discover — already deep
Sparse dynamics, symbolic search, lagged structure, Pareto frontier, regimes, uncertainty, joint parameter refinement, causal hypotheses. Tunable per run.
2. Understand — explain
Turn a world into meaning: plain-language description of each law, the dominant terms, discovered regimes, dependency/causal hypotheses, fit quality, and the assumptions a result is contingent on. Answers "what did it find, and can I trust it?"
3. Use — forecast / simulate / intervene
Run the world forward, forecast beyond the observed window, and ask what-if: change a parameter or input on a schedule and see how the trajectory responds. Scenario objects make interventions first-class and comparable.
4. Compare — compare
Diff two worlds (or two runs): structure, parameters, complexity, and fit — and diff two scenarios of the same world. Model selection made legible.
5. Share — report
Generate a self-contained HTML report from a .lsworld: rendered equations,
fit and Pareto candidates, regime timeline, uncertainty bands, and inline SVG
trajectory + phase-portrait charts. No server, no external assets — a single
file a colleague can open.
6. Organize — library
A local world library: register, tag, search, and describe .lsworld bundles
so a workspace of discoveries stays navigable.
7. Explore — Studio
The visual product: a discovery canvas, equation explorer, regime timeline, uncertainty lens, and world lab, driven by the shared TypeScript packages.
Planned product family: Quantitative research
The proposed quant family makes LawSynth a local-first research workbench for market-data preparation, stochastic processes, calibration, derivatives pricing, portfolio and market risk, generative stress testing, backtesting, market microstructure, and independent model validation. Every experiment is bound to its data, code, configuration, model, seed, and result hashes.
The family architecture, project boundaries, dependency order, and release gates
live in
docs/roadmap/quant-research-platform.md.
First vertical: Quant Diffusion Stress Engine
The proposed G-SMSE finance pack combines seeded classical SDE and jump-diffusion scenarios, optional conditional deep diffusion generators, portfolio valuation, and governed tail-risk reports. Classical baselines remain readable and reproducible. A learned generator can contribute scenarios only after held-out statistical, privacy, and downstream risk tests pass.
The current product does not ship the full engine. LawSynth has SDE discovery and
a narrow Euler-Maruyama helper, but lacks correlated Heston paths, jump processes,
portfolio valuation, learned diffusion training, and the product surfaces needed
for an end-to-end stress test. The complete boundary and release gate live in
docs/roadmap/quant-diffusion-stress-engine.md.
Principles
- Interpretable first. If a user can't read and reason about the result, it isn't a LawSynth result.
- Local-first & reproducible. Same inputs → same world, offline, forever.
- Honest about uncertainty. Boundaries, assumptions, and confidence are surfaced, never hidden.
- Composable. CLI, SDK, Studio, and services all operate on the same
validated World IR and
.lsworldbundles.
Roadmap (product)
- Now:
report,explain,compare,forecast/intervene,libraryacross CLI + SDK; Studio discovery/equation/regime screens. - Next: scenario boards, model-selection views, notebook-native rich display, world templates.
- Later: self-hosted collaborative workspaces and an open plugin registry.
- Proposed domain releases: quant foundation and classical pricing/risk precede generative stress testing; each starts only after a design partner, licensed evaluation data, boundary specs, independent references, and model-risk thresholds are approved.