dsh-ultramath: A Multi-Agent Math Modeling Contest Solver for DeepSeek Harness

What Does It Do?

UltraMath is a DSH plugin for DeepSeek Harness that packages award-winning mathematical modeling competition methodology into native agent presets and skill packs. It solves the tedious and inefficient process from problem reading to paper review in math modeling contests, letting you run the full pipeline of "read problem → framework → derivation → coding → verification → paper → review" directly inside the harness. Core capabilities include 5 role presets, a 33-document model library, independent verification, three-seat blind review, number freezing, and academic integrity gates, all fully self-contained with no external dependencies.

Core Features

• 5 role presets: the UltraMath main controller covers the full pipeline, while mathematician, engineer, writer, and reviewer presets allow single-stage entry, idempotently synced to local on startup.
• 33-document model library plus 93-document abstract methodology: loaded on demand via trigger-word index and method-selection decision tree.
• Full pipeline with quality gates: Phase 0 framework → 1 derivation → 2 coding → 2.5 independent verification → 3 paper → 4 review, each with gates and rollback on blocking issues.
• Three-seat independent blind review with structured scorecard: 3 judges with isolated contexts score independently; disagreements over 20 points go to evidence arbitration, not averaging.
• Number freezing and propagation detection: frozen_numbers.json makes every number in the paper traceable to CSV/script output, with stale-reference detection after code changes.
• Academic integrity 7-category gates: plagiarism, data fabrication, citation forgery, etc., any trigger causes one-vote veto.

How to Use This Plugin?

After enabling it in DSH, create a new session, select the "UltraMath" main controller, and paste the problem statement with data attachments to run the full pipeline automatically. For a single step, specify in natural language, e.g., "Only do modeling and solving for problem 1, don't write the paper yet" or "Paper is done, just run review convergence." For human-in-the-loop control, say "manual mode" or "friendly" to pause at each key decision point with numbered options. For method selection, ask "What model should I use for this problem?" and the plugin recommends and loads the relevant model library document via the trigger-word index and decision tree. To resume mid-run, place the existing solving/ directory in the workspace and say "continue solving" to pick up from the checkpoint.