Foreword

In the DeepSeek Harness ecosystem, existing plugins already cover functions such as literature discovery, reading notes, general RSS subscriptions, and ruling on scientific claims. The dsh-frontier-repro plugin does not duplicate these surface-level features. Instead, it focuses on the “missing path” from curated raw signals to versioned evidence bundles and then to auditable reproduction decisions. It acts as an evidence-first frontier AI radar and reproduction gate, helping developers filter noise and construct concrete reproduction paths.

Plugin Positioning

The plugin is named dsh-frontier-repro, maintained by JimChen-g, classified under workflow, and licensed under MIT. It aims to turn major signals into versioned evidence bundles and verified reproduction decisions.

Core Features

  1. Cross-source timeline: It integrates timelines across arXiv, laboratory releases, official models, blogs, and X signals.
  2. Identity provenance and ranking: It provides explicit identity provenance and explainable ranking, rather than opaque quality labels.
  3. Evidence matrices: It provides mode-specific evidence matrices for exact, scaled, and behavioral reproduction.
  4. Claim-level protocols: It supports claim-level protocols, including execution, partial reproduction, and from-scratch reproduction.
  5. Resource tracking: It records multi-attempt resources (GPU/CPU/VRAM/time/cost/data).
  6. Dependency graph: It provides a deterministic dependency graph.

Installation and Enablement

The plugin is installed as an npm package. Before use, ensure the Node.js engine version meets the requirement: node ^22.19.0 || >=24.0.0.

Typical Usage

The plugin provides a set of command-line tools for handling frontier release evidence and reproduction workflows:

  • frontier_repro_status: Check the health status of sources, corpora, credentials, and persisted sources.
  • frontier_repro_collect: Refresh curated sources, persist, deduplicate, and rank signals.
  • frontier_repro_events: List versioned cross-source evidence bundles and substantive monitoring changes.
  • frontier_repro_bundle: Inspect an evidence bundle, predecessor chain, source records, claims, and all attempts.
  • frontier_repro_get: Retrieve full provenance, artifacts, evaluations, and execution history.
  • frontier_repro_assess: Run evidence gating for exact/scaled/behavioral reproduction.
  • frontier_repro_assess_claims: Perform claim-level gating for execution/partial/from-scratch reproduction modes and exact/scaled/toy equivalence.
  • frontier_repro_graph: Generate the deterministic source/artifact/requirement/execution dependency graph.
  • frontier_repro_record_result: Persist executed commands, artifacts, metrics, deviations, and adjudications.
  • frontier_repro_record_attempt: Store a claim-level attempt, resources, verifiers, and honest outcomes.
  • frontier_repro_manifest: Create a canonical reproduction handoff manifest with SHA-256 integrity.
  • frontier_repro_trackio_scaffold: Export a local-first Trackio logging scaffold.

Applicable Scenarios and Notes

  • This plugin is not a general news reader, arXiv search engine, or paper summarizer.
  • It filters company/personnel announcements, placeholder titles, sources lacking date records, and personal-life posts.
  • Sources can declare category allow/deny lists.
  • It only inspects recent Hugging Face model cards to discover links; it does not archive complete card text.
  • arXiv records retain stable paper IDs and observed immutable vN.
  • GitHub repositories and Hugging Face models/datasets are resolved to full SHAs.
  • The plugin runs with the current DSH process permissions; review the source code and license before installation.

Closing

By converting major signals into auditable evidence bundles, the plugin adds frontier AI research radar and reproduction gating capabilities to the DeepSeek Harness ecosystem. More details can be found on its GitHub page or in the community directory.