Preface¶

In the DSH plugin-based workflow, tasks such as mathematical modeling, prediction, optimization, estimation, simulation, and mechanism analysis often encounter a common issue: code runs successfully, the optimizer returns a solution, and the model seems reasonable—yet none of this guarantees the conclusion is correct. yohanchen1/MathModelingAgent addresses these open-ended tasks by registering key conclusions as Claims and verifying them through Obligations and Evidence; when evidence is insufficient, it maintains an INCONCLUSIVE status rather than forcibly asserting a conclusion.

What This Is¶

MathModelingAgent is a DSH plugin designed for open-ended mathematical modeling, prediction, optimization, estimation, simulation, mechanism analysis, and mathematical modeling competitions. The GitHub repository is yohanchen1/MathModelingAgent, owned by yohanchen1; in DSH uninstall commands, the package name is dsh-math-modeling-agent. This article pins the installation command to v0.3.1, and the license is MIT.

It provides two Skills:

• math-modeling-agent: For building and advancing models.
• math-modeling-audit: For independently auditing existing papers, models, or code, issuing PASS / FAIL / INCONCLUSIVE verdicts item by item without modifying the original text.

Core Features¶

Claim, Obligation, and Evidence¶

Key conclusions are registered as Claims. Verification around Claims proceeds through Obligations and Evidence. There are three verdict statuses: PASS, FAIL, and INCONCLUSIVE.

The INCONCLUSIVE status must be preserved deliberately: it indicates that evidence is insufficient to determine a conclusion, and it cannot be escalated to PASS merely because it seems plausible.

Optional Tools¶

The plugin offers pluggable tool capabilities and does not treat tool outputs directly as final conclusions.

• Python: Optional, used for numerical computation, data analysis, optimization, simulation, plotting, and independent recalculation.
• Lean: Optional, used for formal verification. Lean is not installed automatically; proving a formal proposition does not equate to proving a real-world claim.
• Wolfram: Optional, used for symbolic computation, analytical derivation, and identity verification.
• Literature Research: Supported; private raw data is excluded from retrieval.

Missing tools are never disguised as verification success. When tools are absent, the plugin records the absence, downgrades the evidence level, reduces conclusion strength, and preserves unmet obligations.

Resumable Execution¶

Resumable execution is supported, with the default run directory:

math-modeling-runs/<task-id>/

MCM / ICM Final Review¶

For MCM / ICM-type problems, a final review framework is provided. It must be clarified here: this is not a COMAP official scoring sheet and does not equate to official scoring.

Installation and Activation¶

First, install the plugin:

dsh plugin --profile web add github:yohanchen1/MathModelingAgent#v0.3.1

After installation, check the configuration:

dsh --profile web --dump-config

You should see the dsh-math-modeling-agent-skills line.

Restart to load the plugin:

dsh web

If using the npm source channel, do not replace this with npm install directly, as it will install to the wrong location.

Typical Usage¶

Starting Modeling¶

Simply describe the task. Example:

Build a model for this mathematical modeling problem, first analyzing the problem statement and data.
Any conclusions claiming "optimal," "significant," or "generalizable" must be supported by corresponding evidence; do not force conclusions when evidence is insufficient.

Independent Audit¶

Provide an existing paper, model, or code and request item-by-item verdicts:

Independently audit this result, issuing PASS / FAIL / INCONCLUSIVE item by item, and do not help me modify the original text.

Uninstallation¶

Uninstall command:

dsh plugin --profile web remove dsh-math-modeling-agent

Then restart the current DSH host.

Suitable Scenarios and Notes¶

Suitable for the following uses:

• Handling open-ended mathematical modeling, prediction, optimization, estimation, simulation, or mechanism analysis.
• Needing to separate key conclusions such as “optimal,” “significant,” or “generalizable” from evidence.
• Needing to independently audit existing models, papers, or code while preserving PASS / FAIL / INCONCLUSIVE statuses.
• Using MCM / ICM-type problems with a desire for self-checking via a non-official final review framework.

Usage notes:

• INCONCLUSIVE cannot be escalated to PASS merely because it seems plausible.
• ATTEMPT can never jump directly to SOLVED; SOLVED requires a frozen scope, all mandatory obligations passing, key adversarial checks passing, complete reproducible materials, and declared limitations.
• Missing tools are never disguised as verification success; they only lower the evidence level and conclusion strength.
• The MCM / ICM final review is a non-COMAP official scoring sheet.
• DSH plugins run with the permissions of the current dsh process; check the source code and license before installation.
• Community directory pages are not official app stores; links are only for viewing plugin information and the GitHub repository.

Links¶

• GitHub: https://github.com/yohanchen1/MathModelingAgent
• Directory page: https://www.skillhub.cn/plugins/yohanchen1/MathModelingAgent