Reaper Eagle Forge ML: AMD ROCm Readiness Auditor

Machine Learning teams evaluating a move away from CUDA face a trust problem as much as a technical one: migration benchmark claims are often unreproducible, incompletely instrumented, or asserted with no evidence attached. Reaper Eagle Forge ML audits GitHub repositories for CUDA/NVIDIA lock-in and scores them across four independent axes: portability, benchmark integrity, evidence completeness, and claim discipline. It is not a CUDA-to-ROCm code transpiler and not a raw speed leaderboard. Every claim Forge makes is routed through a claim ledger and, where hardware evidence is involved, backed by a real capsule captured on AMD Instinct MI300X hardware, hash-verified with SHA-256, and clearly labeled as replay rather than live. The permanent deployment has no GPU and honestly reports that, rather than simulating access it does not have. Forge's scanner is deliberately deterministic, not an LLM agent: reproducible pattern matching over untrusted repository input, with zero code execution and zero risk of an inconsistent finding across runs. Semantic, agent-based analysis is scoped as future work, not folded into the current claim. Forge does not ask judges to trust benchmark claims. It shows the code path, the hardware path, the evidence path, and the uncertainty. Disclaimer: The synthetic demo repository used to showcase findings (intentionally seeded with CUDA-lock-in and benchmark-discipline issues) was authored with AI assistance; the deployed scanner itself is fully deterministic and does not call any LLM at runtime.