Research · Advanced
Distilling neural circuits into verified code
Convert extracted transformer subcircuits into human-readable Python algorithms with formal checks
- When you askSubmit circuit artifact
- ThenRun synthesis and solver checks
- ResultPublish verified code and proof
Automate the analysis of pruned transformer subcircuits by synthesizing equivalent symbolic algorithms. When a circuit artifact is submitted, the bot runs domain-bounded surrogate fitting, searches algorithmic motifs, checks equivalence via a satisfiability solver, and reports matching code or counterexamples.
Prompt to start with
BotBentoRecommended
When a new circuit graph is provided, execute the distillation suite via the terminal to synthesize candidate Python representations. Run bounded solver checks against the target domain. If a counterexample or proof is produced, write the findings and verified algorithm into notes. Always ask for confirmation before pushing changes or updating benchmark repositories on GitHub.