An interactive web laboratory exploring how AI proves mathematical theorems through neural proof search, tactic prediction, SAT solving, and the Curry-Howard correspondence.
- 🌳 Proof Tree Explorer — Visualize proof search: goals, tactics, branching, backtracking
- 🧠 Neural Tactic Predictor — Train a mini neural network to predict proof tactics (AlphaProof-style)
- ⚡ SAT→Proof Visualizer — DPLL/CDCL algorithm with conflict-driven clause learning
- λ Curry-Howard Playground — Propositions-as-types: construct proofs by building λ-terms
- 📊 Difficulty Estimator — Predict theorem difficulty from structural features
- 🏟️ Verification Arena — Race neural vs classical proof search on benchmarks
- MIT Technology Review named mechanistic interpretability a 2026 Breakthrough Technology
- Google DeepMind's AlphaProof achieved IMO silver-medal performance in Lean4
- Lean4 is becoming the competitive edge in AI safety verification
- Neural theorem proving combines LLMs with formal verification for trustworthy AI
- Anthropic aims to "reliably detect AI problems by 2027" via interpretability + formal methods
- Zero dependencies — single HTML file
- In-browser neural network training (vanilla JS gradient descent)
- Canvas-based proof tree and SAT graph rendering
- Dark theme, scientific aesthetic
MIT — Built with Claude Code | middesurya