Introduction

When conducting long-range research in DeepSeek Harness (DSH), a common issue is that proposing hypotheses, exploring methods, verifying conclusions, and writing implementations tend to get mixed into the same context, making it difficult to distinguish which conclusions have been verified and which are merely generated results.

dsh-rigorquant provides a set of agent presets and bundled skills tailored for this type of empirical/computational mathematics research. It organizes a single DSH session into a research process that features boundaries, division of labor, and pre-implementation verification.

What is this

dsh-rigorquant is a DSH plugin, with the repository address:

https://github.com/linxichen/dsh-rigorquant

Its description is:

RigorQuant for DeepSeek Harness: session-scoped unattended, context-isolated
multi-agent research for empirical/computational mathematics with a four-part
pre-implementation check battery and a jacobian/Lean escalation lane.

The license is MIT.

Simply put, it turns a DSH session into a context-isolated multi-agent research lab: parallel exploration, independent re-derivation, adversarial checking, meta-validation, escalation paths, and activity panels all collaborate within the same session.

Session Boundaries

The term “unattended” here is explicitly limited to:

  • Only unattended within a single live session.
  • The goal is released after crossing the session boundary.
  • It is re-activated after one round of manual continue input.
  • It does not autonomously continue across reboots.

Therefore, it is suitable for advancing long research tasks within a single session and is not intended for use as an automated research system that spans processes or reboots.

Multi-Role Research Process

The core of dsh-rigorquant is splitting research into multiple roles, separating production from checking.

Verified process points include:

  • Parallel explorers propose candidate methods in a blank context.
  • The ground-truth track re-derives analytic closed forms, invariants, and bounds, using two independent subagent_ground_truth calls.
  • The Adversary only eliminates paths via counterexamples.
  • The Meta-validator rq_check.py checks evidence and rejects PASS when evidence is missing.

The README calls the loop the five moves and names them:

  1. Promise
  2. Fan out
  3. Ground-truth it
  4. Attack it

The verified materials truncate here, and the fifth action is not elaborated upon.

The key constraint of this process is: producers do not check their own work; a path is only eliminated when it encounters a specific counterexample.

Pre-implementation Checks

Before numerical implementation, the plugin runs a four-part check battery.

The verified materials do not list the specific names of the four checks, but confirm that it runs before numerical implementation and is validated by the meta-validator rq_check.py.

Key behaviors of rq_check.py:

  • It reads the audit record, not study.json.
  • It rejects PASS if evidence is missing.
  • Research results cannot endorse themselves.

Randomness and Escalation Paths

For stochastic work, the plugin uses:

fixed-seed + LLN conventions

It also provides an opt-in jacobian MCP escalation lane. Lean is used as a manual external lane.

This means the plugin can take escalation paths on certain critical judgments rather than staying at single-turn reasoning or single generation results.

Activity Panel

The plugin comes with a live activity panel, accessed via:

rq-activity

It only observes the current RigorQuant session, displaying the running status of the current session.

The verified boundaries are:

  • Pure observation, does not change tools, routes, or models.
  • It reads events already published by the core.
  • It provides a JSON snapshot + portraits.

In other words, the activity panel is a read-only view used to observe the research process within the current session.

Installation and Activation

The verified materials do not provide official installation commands that can be copied directly.

package.json contains this information:

name: dsh-rigorquant
version: 0.4.0
license: MIT
bin: dsh-rigorquant -> ./install.sh

The repository file list includes:

cordis.patch.yml
dsh/
agent-presets/
env/
mcp/
docs/
tests/
install.sh
README.md
README.zh-CN.md
LICENSE

It also declares a DSH bundle patch:

dsh.bundle.patch: ./cordis.patch.yml

If you are preparing to install, do not concatenate installation commands from the plugin name. First, check:

  • install.sh
  • mcp/
  • agent-presets/
  • dsh/
  • docs/
  • LICENSE

Confirm the behavior, dependencies, and permission boundaries of these files before enabling.

Typical Usage

After enabling the plugin in a DSH live session, you can hand an empirical/computational mathematics problem to it.

During execution, verified result branches include:

  • PASS: auto-implement and proceed.
  • BLOCKED: after 3 rounds of the same gap, returns the strongest derivation + exact gap.
  • BUDGET: after 5 rounds, checkpoint + report.

When using it, you need to remember the session boundaries:

一次 live session 内：
    可以按流程推进。

跨过 session 边界：
    目标解除。

人工输入一轮：
    continue

结果：
    目标重新激活。

If you see BLOCKED, focus on the strongest derivation and exact gap it returns; if you see BUDGET, focus on the checkpoint and report.

Use Cases and Notes

Suitable for:

  • Long empirical/computational mathematics research that needs to be done within a single DSH session.
  • Verification needs to be completed before numerical implementation.
  • A combination process of parallel exploration, independent re-derivation, adversarial checking, and evidence verification.
  • An opt-in jacobian MCP escalation lane.
  • Lean as a manual external lane.
  • rq-activity to observe the current RigorQuant session.

Notes:

  • It is only unattended within one live session and does not run autonomously across reboots.
  • It is not a standard single-turn Q&A plugin, but a combination of roles, skills, and check processes.
  • The plugin runs with the permissions of the current dsh process; check the source code and license before installation.
  • Current verified materials do not provide an official installation command; do not concatenate commands like dsh plugin add based on the name.
  • rq-activity only observes and does not change tools, routes, or models.

Conclusion

The value of dsh-rigorquant is to transform math research in a DSH session from a single generation into a process with boundaries, division of labor, checks, and evidence. It is suitable for empirical/computational mathematics research scenarios that require auditability and pre-implementation checks.

Repository address:

https://github.com/linxichen/dsh-rigorquant