7 projects
llm-feedback-control
A control layer for LLM-in-the-loop reliability: wrap an unreliable generator in a verify-and-refuse feedback loop with a pluggable controller (deterministic checks, a model critic, or compositions) and composable 'circuits'. Keep one exact element in the loop and outputs are checked, corrected, or honestly refused. Zero runtime dependencies; runs with no model at all.
numeric-rule-finder
Discover the conservation/balance rules a set of numbers must satisfy, find where they break, and honest-stop when there is no such structure.
petra-nn
Petri-Net Trained Architecture — formally-verified learnable process intelligence
structural-computing
Declarative structural computation: exact polynomial-time answers to combinatorial questions on structured graphs, with automatic routing to the cheapest correct evaluator.
holant-tools
Decision procedures for matchgate-Holant tractability: polynomial-time SRP solver, Cai-Lu §4 d-admissibility with full augmented-Pfaffian Plücker enumeration across all viable m configurations, Galluccio-Loebl Holant evaluator, MatchgateFamilyDetector, corrected dart-chain intersection numbers, hardness-candidate screening, and a tropical-Holant kernel with polynomial-time optimisation (Hungarian for bipartite, Edmonds blossom for general).
structural-computing-bench
Calibrate the cost models used by structural-computing's router via wall-clock measurements on your machine.
timeseries-formula-finder
Symbolic Form Discovery - Find interpretable mathematical formulas in data
