Neural Proposer & Formal Verifier Lab Neuro-Symbolic Search

Pairing statistical LLM candidate beam generation with strict deterministic proof kernel verification

Candidate Proof Search Tree Ready
Verified (Kernel Check)
Hallucination Caught
Selected Active Path
Pruned / Dead End
7
Tree Nodes Explored
1
Hallucinations Rejected
5
Axiomatically Sound Steps
VERIFIED
Kernel Proof Status

Why Statistical AI Needs a Deterministic Formal Referee

Large Language Models (transformers) operate on statistical token transitions. While remarkably proficient at proposing creative lemma substitutions and guiding proof search heuristics through vast combinatorial spaces, their lack of an internal truth model causes subtle mathematical hallucinations (e.g., falsely assuming non-commutative group elements commute or applying false geometric congruence).

1. Statistical Proposer (Creativity & Guidance)

Scores candidate next steps via learned mathematical priors. Discovers non-obvious auxiliary constructions (like in Google DeepMind's AlphaGeometry) that purely brute-force symbolic search could take centuries to encounter.

2. Deterministic Formal Verifier (Unforgiving Truth)

Executes strict axiomatic unification (Lean, Coq, Isabelle). Instantly rejects invalid leaps with zero tolerance for probabilistic "vibes", ensuring generated proofs meet rigorous mathematical certainty.

Enjoy this tool? Build your own with Super