dsh-rigorquant

English | 简体中文

Unattended-within-a-session, long-running empirical/computational mathematics research for DeepSeek Harness — economics, finance, portfolio construction/optimization, simulation, computational econ/finance.

RigorQuant is an agent preset + bundled skills that turns one DSH session into a context-isolated multi-agent research lab:

• J-Space is used integrally across the root persona, every subagent role, and plan mode as the inference-time cognitive-control layer (workspace gate, ledger, seam refresh, dense inner / clean outer registers).
• Parallel explorers propose candidate methods (subagent, blank context).
• A ground-truth track re-derives the analytic closed forms, invariants, and bounds for simplified cases — twice, by different means (two independent subagent_ground_truth calls).
• An adversary eliminates routes by counterexample only.
• A four-part check battery (closed-form equality, exact invariants, analytic bounds, statistical hardening) runs BEFORE numerical implementation.
• A meta-validator (rq_check.py) refuses a PASS whose evidence is missing: empty stage outputs, an empty derivations/, a registry with no audit-referenced passed route, or deliverables that do not compile. Its evidence checks read the audit record, not study.json — a study may not vouch for itself.
• Fixed-seed + LLN conventions for stochastic work.
• A jacobian MCP escalation lane (opt-in; Lean as a manual external lane) settles proof-critical claims before implementation.
• PASS → auto-implement and proceed; BLOCKED → 3 rounds of the same gap → strongest derivation + exact gap; BUDGET → 5 rounds → checkpoint + report.

The operating pattern adapts Shanmu Jin's Crouzeix-conjecture run (prompt, Lean audit) and Terence Tao's blueprint/equational-theories projects to numerical work. Full design record: docs/architecture.md.

"Unattended", precisely: the framework runs unattended within one live session. Crossing a session boundary disarms the goal; one human turn ("continue") re-arms it. It does not continue autonomously across restarts.

Install

Requires DSH ≥ 0.1.0-rc.7.

One line, everything — the preset, the compute lane, and the plugin (role model router + its Settings card):

npx dsh-rigorquant
# npx dsh-rigorquant --profile tui     # a profile other than web

From a clone — the same install, from your own working tree. A checkout installs itself into the profile, so re-run this after editing dsh/ to refresh the plugin:

git clone https://github.com/linxichen/dsh-rigorquant
cd dsh-rigorquant
./install.sh
# ./install.sh --skill-only     # only the skills, for any preset, no plugin
# ./install.sh --uninstall      # removes everything, plugin included

Plugin only — the same everything, via the ecosystem's bundle path. The package declares a dsh.bundle manifest whose rows include a boot-sync half (rq-preset-sync): on the profile's next start it lands the agent preset into $DSH_HOME/.agent-presets/rigorquant and the compute lane into $DSH_HOME/share/rigorquant/, so dsh plugin add alone yields a working distribution (docs/architecture.md Decision 23):

dsh plugin --profile web add dsh-rigorquant

The sync is idempotent (a byte-identical tree is left untouched; derived state like .venv is never copied or pruned) and keeps same-version local edits to the installed preset — an upgrade replaces shipped files, exactly like re-running ./install.sh. There is no uninstall hook in DSH's plugin CLI, so removal stays explicit (./install.sh --uninstall); if you remove only the plugin, the synced preset keeps working standalone and simply routes nothing.

Start a new DSH session and pick the RigorQuant preset. Then:

rigorquant: derive and validate a method for [problem], simplified cases first, before any numerical implementation.

Compute lane (one-time)

The pinned uv lane lives at $DSH_HOME/share/rigorquant/env, placed there by install.sh or by the plugin's boot-sync row — whichever ran last owns the anchor, and both land identical bytes (see env/README.md). The venv itself is never installed: it is derived state, provisioned lazily inside the anchor by the first uv run --frozen --project <env_lane> (subsequent runs are instant; --frozen honors the committed lockfile). The jacobian escalation lane ships disabled and pinned (jacobian@0.12.0): enable the mcp-jacobian row, and the framework asks for approval before any one-time provisioning (npx -y jacobian@0.12.0 upgrade, or the Lean toolchain via the skill's scripts/provision-lean.sh). See mcp/jacobian.md.

Role-routed models (rq-model-router)

The bundled plugin routes each RigorQuant role to its own model + reasoning effort, with one fallback per role. Configure it in Settings → Plugins → RigorQuant model routing: the last saved selection persists (settings user layer). Shipped defaults:

On a terminal primary failure (no adapter / HTTP 4xx) the role degrades to its fallback for one forced retry, and recovers on the next success or after 10 minutes. Untagged agents (other presets, workflow workers, forks) are never touched. Requires DSH ≥ 0.1.0-rc.7 (self-registered plugin settings). Design record: docs/architecture.md Decision 16.

Repository layout

package.json                dsh.bundle manifest (dsh plugin add support)
cordis.patch.yml            bundle patch: skills layer + rq-model-router +
                            rq-preset-sync rows
dsh/                        host halves (rq-model-router router + rq-preset-sync
                            boot-sync) and the Plugins-tab card
agent-presets/rigorquant/   preset composition + persona + bundled skills
  skills/rigorquant/        SKILL.md + references/ + scripts/ + schemas/
  .../scripts/rq_check.py   the meta-validator (single canonical copy)
  .../schemas/              study.json + registry.json JSON Schemas, which the
                            validator loads — so schema and checker cannot drift
env/                        pinned uv compute lane (sympy/cvxpy/hypothesis/…)
mcp/jacobian.md             escalation lane wiring
docs/architecture.md        grilled decision record + sources
tests/                      the validator's test suite (see Testing below)
studies/                    one study folder per task (Mode B; a checkout's own
                            live studies — not shipped in the npm bundle)

Testing

The validator has a test suite, and its centrepiece is a forged study — empty derivations, empty stage outputs, a one-line adversary report, and a paper whose body reads "This paper says nothing." It must FAIL. A framework whose honesty gate is not itself tested is a framework that certifies whatever it is handed.

uv sync --frozen --project env
uv run --frozen --project env python -m pytest tests/ -q

tests/test_repo_consistency.py covers the other half: one validator, one schema, documented commands that resolve, and layout blocks that match the filesystem. That is the defect class this repository actually produces.

What a green validator means: nothing declared is missing, and the deliverables build. It does not mean the mathematics is right — that stays with the check battery, the independent ground-truth track, and the adversary.

Studies

A study is one self-contained rigorquant task with an identical folder structure everywhere: durable deliverables at the study root (study.json, STUDY.md, registry.json, journal.md, derivations/, audits/, artifacts/) are meant to be committed; all scratch lives in a gitignored interim/. Two modes, implied by location:

• One study per repo — study.json at the repo root.
• Multiple studies per repo — studies/<slug>/study.json; the roster is studies/*/study.json.

Intake detects an existing study and continues it silently; a new study asks one question (mode + slug) and never asks again. See docs/architecture.md §12.

Publishing

This repo is a community DSH plugin distribution (bundle + preset + skill form): it declares a dsh.bundle manifest in package.json, is tagged dsh-plugin, and is discoverable by the ecosystem's topic-based indexes — see dsh-find-plugins and the awesome-deepseek-harness list for the conventions.

MIT License.