caelum0x 20750c64cd docs+P10: competitive analysis, expansion-v2 roadmap, distributed discovery
Competitive intelligence (code-level study of PySINDy, PySR/SymbolicRegression.jl,
Operon, SciML/DataDrivenDiffEq+MTK, gplearn/DEAP, AI-Feynman, + SRBench/
forecasting/MLOps landscape):
- docs/research/competitive-analysis.md — market map, feature matrix (LawSynth vs
  field), and prioritized adopt/court/differentiate recommendations.
- docs/roadmap/expansion-v2.md — a ~10k-file expansion plan (v2-A engine breadth,
  v2-B ecosystem/adoption, v2-C platform/distributed, v2-D domain packs, v2-E
  docs), sequenced by ROI.

P10 distributed discovery (crates/lawsynth-discovery/distributed.rs):
discover_partitioned() parallelizes feature-library evaluation across threads,
BIT-IDENTICAL to single-node (equivalence test at partitions {1,2,3,7,64}),
1.87x at 8 threads; RSS reduction kept single-node for float exactness. The
'distributed AND reproducible' guarantee. 41 tests pass.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-08-18 02:20:12 +03:00
..
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00
2026-08-16 22:04:11 +03:00

Research notes

This directory records how to make a LawSynth result inspectable and repeatable. It describes the methods that are implemented in the current Rust workspace: validated numeric datasets, deterministic differentiation and feature construction, sparse or bounded symbolic candidate search, deterministic scoring, World IR bundles, and numerical simulation.

These notes are not a claim that a selected equation is uniquely true or causal. A result is scientific evidence only when its data provenance, preprocessing choices, model class, held-out evaluation, and known failures are retained with it. Start with methodology, then use reproducibility when preparing a result for another person to rerun.