Sahir619/fable-method

The idea is to reverse-engineer and codify how a particular agentic model approached tasks — breaking it into a think, act, prove loop — then make that pattern model-agnostic and testable. The eval component is what keeps this from being a blog post: there is a harness that checks whether your model's outputs actually follow the workflow, not just whether they look like they do. That said, the value depends almost entirely on how rigorous the eval is and how well the distilled pattern generalizes beyond the original model's training data. Reservation: this is a methodology repo, not a library you install. Adoption requires reading, adapting, and trusting that the abstracted workflow is accurate — and the repo is too new to have external validation. Useful as a starting point for teams who want structured agentic reasoning but not as a drop-in solution. -> Best for: AI engineer or technical PM building structured agentic reasoning into an existing product.