dsh-frontier-repro

Frontier AI releases arrive daily, yet the path from a single arXiv signal or lab blog post to an auditable, claim-level reproduction verdict is often missing: Is the evidence complete? Are commands and artifacts preserved? Are failures and negative results actually recorded? dsh-frontier-repro exists to close that gap.

It aggregates signals from arXiv, OpenAI, Anthropic, Google DeepMind, NVIDIA, AMD, verified blogs, and Hugging Face model repositories into conservative release evidence bundles, each retaining immutable version chains, provenance, capability and license diffs, and explicit missing-evidence markers. At the reproduction stage, the plugin enforces claim-level protocols across execute_existing, partial_reimplementation, and from_scratch_replication modes at exact, scaled, or toy equivalence. A successful verdict requires commands, artifacts, measured metrics, and independent verifier evidence; failures and negative results are persisted just as rigorously. The evidence dependency graph, SHA-256 integrity manifest, and Trackio scaffold export make the full pipeline traceable and handoff-ready.

Built for researchers, ML engineers, and evaluation teams who need an auditable reproduction pipeline. If you want every release claim to pass through a complete evidence gate rather than resting on a summary or an opaque quality label, this plugin belongs in your workflow.