Algorithm Forge

AI-generated code often looks correct and still fails on cases nobody tested. Algorithm Forge is a research lab built around one rule: AI proposes, evidence decides. Meta Muse Spark proposes algorithms. Algorithm Forge runs them in isolation and checks them with an independent, deterministic verifier. Every result is recorded and reproducible. It demonstrates two failure modes that model confidence alone cannot reveal: 1. Valid is not optimal. In scheduling, a model's solution can be feasible but provably suboptimal. Algorithm Forge verifies a candidate makespan of 4 against a mathematically proven optimum of 3. 2. Plausible is not correct. In a recorded Tricky Two-Sum experiment, independent verification found 1,175 failures across 51,831 test cases and kept the smallest counterexample. A higher-reasoning configuration of the same model recorded zero failures on the same checker. The platform includes an Experiment Studio, experiment history, scientific dashboards, scheduling verification, inspectable counterexamples and reproducible evidence. Algorithm Forge is a research lab, not a product. Its contribution is the method: every claim a model makes is checked by a trusted verifier, and "valid" is never confused with "optimal". The same approach applies to any team that needs to decide whether AI-generated code can be trusted.