graphopt — the research stack

graphopt is a categorical optimizer for neural-network computation graphs — and a machine that writes its own rewrite rules, certifies them, and learns to play the game they create. Programs, rules, and derivations are all values in one calculus (n-morphisms), so a minted rule re-enters the bank as a legal move and discovery closes into a self-modifying game under a certifying referee.

• How graphopt works — syntax → structure → search → certificate.
• Evaluation: the fourth dimension — describe, rank, measure — never decide equivalence.

Cross-cutting concepts

• The three costs — dag_cost guesses, fixtures measure, the verifier proves.
• The game — boards, players, episodes, and the learned-policy loop.
• The gauntlet — what a minted law must survive to ship.

Learn by doing

The tutorial track walks install → first search() → first minted law → the scorecard. The contract surface is the frozen API reference; upstream, the repository's own docs live at handl-lang/graphopt/docs/.