You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
RM is a network-free simulator with strict semantics enforcement (block_same_complex_binding=true by default; equivalent to NFsim's -bscb on both binding and unbinding sides). Cleanroom C++17 implementation, MIT licensed.
The reason this might be useful to NFsim: the test infrastructure under tests/ is fully portable. Three corpora, 177 models, all with vendored reference trajectories:
feature_coverage/ — 77 small, fast models, one BNGL feature per model, plus combo_* interactions and 20 edg_* edge-case stress probes specifically designed to break things. NFsim 20-rep reference + BNG2 ODE cross-check per model. ~3 min full run.
corpus/ — 71 real-world'ish rule-based models. NFsim 100-rep ensemble references with hand-rolled Gillespie SSA fallback for two cases where NFsim was found to diverge (issue Disjoint reactant patterns under-count embeddings with multiple matching molecules #62 and a related ring case — both since fixed in your tree). ~15 min smoke / ~3 h full.
nfsim_basicmodels/ — 29-model adapted version of NFsim's own validate/basicModels/ suite. ~5 min.
Harness is harness/benchmark_*.py. References + manifests are in-tree. Each suite emits a per-model PASS/FAIL with z-score thresholds (tz_max < T_model). All 177 currently green against RM 3.1.2.
The corpora and harness are MIT-licensed and available in-tree if any of it would be useful upstream.
Heads-up: RuleMonkey 3.1.2 is out — https://github.com/richardposner/RuleMonkey/releases/tag/v3.1.2
RM is a network-free simulator with strict semantics enforcement (
block_same_complex_binding=trueby default; equivalent to NFsim's-bscbon both binding and unbinding sides). Cleanroom C++17 implementation, MIT licensed.The reason this might be useful to NFsim: the test infrastructure under
tests/is fully portable. Three corpora, 177 models, all with vendored reference trajectories:feature_coverage/— 77 small, fast models, one BNGL feature per model, pluscombo_*interactions and 20edg_*edge-case stress probes specifically designed to break things. NFsim 20-rep reference + BNG2 ODE cross-check per model. ~3 min full run.corpus/— 71 real-world'ish rule-based models. NFsim 100-rep ensemble references with hand-rolled Gillespie SSA fallback for two cases where NFsim was found to diverge (issue Disjoint reactant patterns under-count embeddings with multiple matching molecules #62 and a related ring case — both since fixed in your tree). ~15 min smoke / ~3 h full.nfsim_basicmodels/— 29-model adapted version of NFsim's ownvalidate/basicModels/suite. ~5 min.Harness is
harness/benchmark_*.py. References + manifests are in-tree. Each suite emits a per-model PASS/FAIL with z-score thresholds (tz_max < T_model). All 177 currently green against RM 3.1.2.The corpora and harness are MIT-licensed and available in-tree if any of it would be useful upstream.