Add lagh: certified symbolic law discovery - #214
Conversation
lagh returns three-track output: machine-checked exact certificates (exhaustive per-point check under a declared error model, with a stated chance-fit significance bound alpha), labeled empirical conjectures, or explicit abstentions. For this harness the conjecture track always answers; certificate status is carried as estimator metadata (track_/tag_/alpha_log10_). Zero-confident-wrong invariant. Method source: https://github.com/jascal/lagh (Apache-2.0), installed from the stable repo via install.sh per the contribution guidelines.
…avalab/srbench#214) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01KCkX3udh52kzBvzkWEwc77
actions/cache v1/v2 were closed down by GitHub in early 2025; every job now fails at 'Set up job' before running. Bump cache v2->v4, checkout v2->v4, setup-miniconda v2->v3 in the PR-triggered workflow only.
|
Note for reviewers: the |
conda-forge sunset Mambaforge; the latest-release download now 404s and setup-miniconda's own warning says to switch. Miniforge3 ships mamba.
…tion() Same regressor file works under both the dev contract and the 2025 living-benchmark contract (experiment/methods layout can be mirrored on request).
The root environment's pip section transitively installs the retired 'sklearn' package, which now hard-errors; set the documented escape hatch so the baseline env can build.
|
Closing this: the dev branch's CI has accumulated several layers of retired-infrastructure breakage (actions/cache v2, Mambaforge installers, the PyPI sklearn shim), and it doesn't make sense to debug that inside a method submission. The method directory remains available ready-to-use at https://github.com/jascal/lagh/tree/master/srbench-submission/algorithms/lagh (contract-compliant for both the dev and srbench_2025 layouts) — happy to resubmit if the intake pipeline gets refreshed. Thanks for maintaining the benchmark. |
1 participant
What this adds
algorithms/lagh/— a new SR method per the contribution guidelines (metadata.yml, regressor.py, install.sh pulling from the stable source repo, requirements.txt).lagh (github.com/jascal/lagh, Apache-2.0) is a certified symbolic law discoverer: its native output is three-track — a machine-checked exact certificate (every training point within a declared epsilon; a chance-fit significance bound α ≤ |H|·q^h is computed and null-validated), a labeled empirical conjecture, or an explicit abstention. The design goal is a zero-confident-wrong invariant: the system never claims an exact law it cannot certify, and its refusals state why.
Harness mapping
estis sklearn-compatible withmax_time(SIGALRM handled; on timeout an affine-OLS fallback conjecture is stored soevaluate_model.pyalways finds an equation) andrandom_state.model(est, X)returns a sympy-compatible string withx_imapped to the training DataFrame's column names per the guidelines.est.track_,est.tag_,est.alpha_log10_) for anyone who wants to slice results by claimed-exact vs conjectured.Provenance
The method's benchmark record (a registered one-shot read of Gravity-Bench, an honestly-reported LLM-SRBench read, and four real-archive case studies with reproduction commands) is documented in the source repo's README and docs/. No source code is vendored here;
install.shinstalls from the repository per guideline 7.Happy to adjust anything to fit the evaluation setup — thanks for maintaining the benchmark.