Inference token buffer judge: recipe template

Use this recipe to prepare and evaluate a judge for bounded ready/valid RTL repairs. It contains no trained weights, serving endpoint or measured model accuracy.

The seed corpus is pinned to Dataset revision 1: a written contract, original MIT-licensed RTL, an independent native scoreboard and a measured regression with 34,909 checks across 4,111 cycles and zero errors. Executable checks and engineering review provide the proposed training labels.

Build on the seed

1. Read the buffer contract and reproduce the native testbench in the dataset.
2. Prepare independently reviewed correct and faulty implementations. Keep changes and their expected effects traceable.
3. Choose a compatible base model and license.
4. Separate held-out designs and mutations from the training examples.
5. Specify the training configuration and evaluation plan before fitting a model.

The supplied recipe.json pins the starting evidence and records the remaining inputs. README.md explains the workflow. Success on this one buffer does not establish a learned judge’s generalization, physical signoff or complete accelerator performance.