Mathematical Modeling Workflow Entry
This skill is responsible for identifying the competition mode, initializing the project, recording key decisions, and routing subsequent stages. It does not replace the modeling, coding, writing, or verification stages.
Read
../_references/math_modeling_norms.md
when domain standards are needed; when identified as CUMCM 2026,
../_references/CUMCM_2026_RULES.md
must also be read. Competition-specific rules take precedence over general best practices.
Startup Confirmation
First infer known information from the user-provided problem statement, existing files, and
, and confirm only missing items that would change the workflow:
- Competition type and year: if the competition is a domestic contest, national contest, or CUMCM and the year is unspecified, confirm whether it is the current CUMCM 2026 edition.
- Typesetting engine: LaTeX or Typst; default to LaTeX when unspecified.
- Paper language; MCM/ICM/COMAP defaults to English, while other competitions default to Chinese.
- Whether the current phase is pre-contest training or the official contest period.
- Confirm the number of subproblems only when it cannot be determined from the problem statement and would affect initialization.
Do not repeat questions about choices already recorded in
.
CUMCM 2026 Mode
After identifying CUMCM 2026, record the following in
:
yaml
competition_mode: CUMCM_2026
ai_compliance: true
human_review_required: true
restricted_current_problem_search: <true during official contests; false during training>
submission_packaging: true
The following invariants must be observed in this mode:
- AI only assists the work; the team must lead the core modeling and analysis, and all content involving AI participation must be individually reviewed and verified by humans.
- H1–H4 must be genuinely confirmed by the team. The Agent must not enter on the team's behalf, treat silence as consent, or bypass a waiting state.
- During an official contest, do not browse, publish, or discuss content related to the current contest problem on any communication platform, and do not search for answers, approaches, code, or papers related to the current problem. It is permissible to consult public background knowledge needed to solve the problem, official data, academic papers, and software documentation.
- Record every external source actually used in ; record all AI participation in
compliance/AI_USAGE_LOG.md
Initialization Artifacts
Always create or update the following:
- : user choices, mode switches, stage artifacts, risks, and routing.
- : stage checklist and current status.
CUMCM 2026 also requires the following:
text
compliance/
├── CUMCM_2026_CHECKLIST.md
├── AI_USAGE_LOG.md
├── HUMAN_REVIEW_LOG.md
├── SOURCE_LOG.md
├── SUBMISSION_MANIFEST.md
└── private_identity_terms.txt
private_identity_terms.txt
is used only for local anonymity scanning and must not be included in the paper, supporting materials, or submission package. Compliance logs must retain an empty-entry structure, but must not prematurely claim that content has been reviewed, AI has not been used, or checks have passed.
Immediately after initialization, record this Agent's actual participation in workflow planning, problem file inventorying, or rule configuration in
; do not wait until the coding stage to begin logging.
Plan and Directory Structure
In
, specify the actual competition, engine, language, number of subproblems, and mode switches, and assign stages according to the following responsibilities:
| Stage | Skill | Main Artifacts |
|---|
| Problem understanding, candidate models, and human decisions | | ANALYSIS_MODELING_REPORT.md
MODEL_SELECTION_REPORT.md
|
| Programming, baselines, validation, and data visualization | | , , |
| Non-data diagrams | | , , |
| Paper writing | | |
| Scientific, structural, and formatting verification | | |
| CUMCM-specific compliance and packaging | | , supporting material archiving, |
Adjust the files in the generic project skeleton as needed for the actual problem:
text
.
├── plan.md
├── todo.md
├── compliance/
├── reports/
│ ├── ANALYSIS_MODELING_REPORT.md
│ ├── MODEL_SELECTION_REPORT.md
│ ├── RESULTS_REPORT.md
│ ├── DRAWIO_REPORT.md
│ └── VERIFY_REPORT.md
├── code/
│ ├── run_all.py
│ ├── problem1.py
│ ├── ...
│ ├── utils.py
│ ├── requirements.txt
│ └── outputs/
├── results/
├── figures/
├── paper/
└── support_materials/
Non-CUMCM projects do not need to create unrelated compliance directories or invoke
.
To-Do List and Human Gates
For CUMCM 2026, use the following format in
:
markdown
# To-Do Items
- [ ] 1. Parse the problem statement and attachments
- [ ] H1. Team confirms understanding of the problem
- [ ] 2. Candidate model design, independent review, and comparison
- [ ] H2. Team confirms the core modeling approach
- [ ] 3. Program implementation, baselines, and result validation
- [ ] H3. Team verifies the core results
- [ ] 4. Non-data diagrams
- [ ] 5. Paper writing
- [ ] H4. Team reviews the paper and AI-assisted content
- [ ] 6. Scientific, numerical, structural, and formatting verification
- [ ] 7. CUMCM compliance review and submission packaging
For non-CUMCM modes, retain the standard 2→3→4→5→6 workflow. After completing each stage or human gate, update
immediately. When awaiting human confirmation, record an explicit
status and do not proceed past that gate.
Stage Boundaries
- is responsible for actual computation results and data visualization; handles only meaningful flow, structural, and conceptual diagrams.
- may cite only numerical values traceable to result files or chart data and must not re-estimate them.
- When finds substantive upstream errors, route back to the corresponding stage according to the error type; do not remodel during the verification stage.
- is used only in CUMCM mode and is responsible for official formatting, anonymization, AI details, supporting materials, and size limits.