Introduction

The core philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” For mathematical research scenarios, dsh-frontier-math is a frontier mathematical research workflow plugin for DSH. It does not promise to “automatically solve arbitrary open problems.” Instead, it increases the probability of producing genuinely novel and verifiable results through rigorous topic selection, open-status verification, claim dependencies, and independent verification gates.

The plugin is maintained by inoricon1 and is released under the MIT License. It aims to help researchers establish a structured research process when facing complex mathematical problems, providing end-to-end tooling from project initiation to final certificate generation.

Core Features

The plugin provides four core tools, corresponding to different stages of the research process:

  • frontier_math_project: Create, score, view, and advance research projects.

    • A problem must pass the following hard gates before it can be initiated:
      • Open-status evidence ≥ 3/5
      • Statement clarity ≥ 4/5
      • Independent verification path ≥ 3/5
      • Existing-work risk ≤ 2/5
      • Overall opportunity score ≥ 3.0

  • frontier_math_prior_art: Verify whether the problem has already been solved.

    • A complete audit must cover: exact statements or distinctive fragments, equivalent formulations and duals, the current status of original papers and expert pages, recent preprints and follow-up results, and searches for proofs and counterexamples.
    • Requirements: at least two independent domains, two source categories, two evidence families, and one authoritative source.
    • If a credible solved status, refutation, or status conflict is found, the project is automatically blocked.

  • frontier_math_claim: Record conjectures, empirical findings, derivations, counterexamples, and theorems, and maintain an acyclic dependency graph. If a dependency has not been verified, the conclusion cannot be marked as verified.

  • frontier_math_certificate: Record certificates from Lean, exact computation, SAT/SMT, interval algorithms, literature, and expert review. The first version only records independent verification results and does not execute arbitrary commands.

Installation and Activation

Run the following commands in the plugin directory to install and verify:

npm run verify
dsh plugin --profile web add link:.

After installation is complete, restart the DSH Web service. You can check whether the configuration is active with the following command:

dsh --profile web --dump-config

The DSH Web browser research status panel is available at:

http://127.0.0.1:3080/frontier-math

Typical Usage

The recommended workflow includes the following steps:

  1. Ask the model to create a project and provide item-by-item topic-selection scores.
  2. Complete the five categories of prior-work searches; classify each source individually.
  3. For any “solved/partially solved/equivalent proposition” source, write an equivalence explanation.
  4. Proceed to the research phase only when the audit result is provisionally_open.
  5. Establish multiple distinct proof, counterexample, and construction routes.
  6. Break core steps down into a claim dependency graph and preserve failed routes.
  7. Use Lean, exact programs, or expert review to generate independent certificates.
  8. Before packaging results, trigger recheck and redo the five categories of searches.
  9. After the automated gates pass, the state moves to ready_for_expert_review; final confirmation is still made by domain experts.

For model configuration, it is recommended to use deepseek-v4-pro, set the reasoning level to max, enable DSH native tool calling, and preserve the interleaved thinking state.

Applicable Scenarios and Notes

  • Data Storage: Projects are stored by default at %DSH_HOME%/frontier-math; each project is a JSON file with a schema version. Writes use a temporary file in the same directory followed by atomic replacement; the Web API is read-only.
  • Automated Search Blind Spots: Automated search has coverage blind spots and cannot replace MathSciNet, zbMATH, or specialized databases.
  • Source Retention: Source classification is performed by the model, but original links and rationale must be retained for manual inspection.
  • Score Meaning: Scores are for ranking only; value judgments and mathematical equivalence require expert involvement.
  • Certificate Generation: The first version does not directly execute Lean or other verifiers; it only records certificates formed by external runs.
  • Concurrency Limits: JSON storage is suitable for personal research portfolios, not for large multi-user concurrent deployments.

Conclusion

dsh-frontier-math provides a rigorous process framework, suitable for developers or researchers who want to conduct structured mathematical research in DSH. For more details and source code, refer to its GitHub repository and skill directory page.