Problem Intake
文件关系全貌请见 [[FILE_RELATIONSHIP_MAP]] · 上游: [[skills/mm-start-contest-v2/SKILL|V2 总控]] · 下游: [[skills/mm-model-strategy/SKILL|Phase 2 Strategy]] · 共享规范: [[skills/_references/SKILL|_references]]
Load First
Read:
../_references/v2_pipeline_contract.md../_references/codex_subagent_protocol.md../_references/rag_usage_contract.md../_references/source_quality_policy.md../_references/problem_type_router.md
Required Outputs
Create or update:
PROBLEM_BRIEF.mdDATA_AUDIT.mdWORKFLOW_STATE.mdreports/AGENT_RUNS.mdreports/INTAKE_GATE.md
Procedure
- Inventory all provided problem statements, PDFs, DOCX files, spreadsheets, CSV files, images, and downloaded resources.
- Extract the exact contest title, background, top-level questions, attachment descriptions, submission requirements, and known constraints.
- Treat only explicitly numbered top-level questions as
ques1, ques2, etc. Do not inflate small details into fake subproblems.
- If local RAG is available, query
cumcm_problems or mcm_icm_problems for similar task wording, attachment patterns, and hidden scoring cues. Record only sourced hits using ../_references/rag_usage_contract.md and ../_references/source_quality_policy.md; core problem facts should come from the current files or S official sources, not unsourced memory.
- Audit each data file for rows, columns, field meaning, units, missing values, abnormal values, duplicate records, encodings, and whether it can support each subproblem.
- Record ambiguities and risks in
WORKFLOW_STATE.md; mark any unreadable or semantically unclear data as HIGH risk.
- If using Codex subagents, run
problem-analyst and data-auditor independently, then summarize their outputs into reports/AGENT_RUNS.md.
PROBLEM_BRIEF.md Structure
# Problem Brief
## Contest And Task
## Source Files
## Top-Level Questions
## Inputs And Outputs
## Constraints And Evaluation
## Dependencies Between Questions
## Ambiguities And Risks
## Next Modeling Needs
## Local RAG Evidence Used
DATA_AUDIT.md Structure
# Data Audit
## File Inventory
## Table And Field Summary
## Missing / Abnormal / Duplicate Data
## Unit And Encoding Issues
## Usable Variables By Question
## Derived Variables Needed
## Blocking Risks
Gate
Write reports/INTAKE_GATE.md with PASS, CONDITIONAL_PASS, or FAIL.
Do not proceed to modeling on FAIL. On CONDITIONAL_PASS, list exactly what the model strategy stage must handle.
1---2name: mm-problem-intake3description: 数学建模竞赛 V2 题面与数据建档阶段。用于读取赛题、附件和资源，拆解子问题，审计数据风险，建立可恢复上下文，并为后续建模与代码阶段提供结构化输入。4---56# Problem Intake78> 文件关系全貌请见 [[FILE_RELATIONSHIP_MAP]] · 上游: [[skills/mm-start-contest-v2/SKILL|V2 总控]] · 下游: [[skills/mm-model-strategy/SKILL|Phase 2 Strategy]] · 共享规范: [[skills/_references/SKILL|_references]]910## Load First1112Read:1314- `../_references/v2_pipeline_contract.md`15- `../_references/codex_subagent_protocol.md`16- `../_references/rag_usage_contract.md`17- `../_references/source_quality_policy.md`18- `../_references/problem_type_router.md`1920## Required Outputs2122Create or update:2324- `PROBLEM_BRIEF.md`25- `DATA_AUDIT.md`26- `WORKFLOW_STATE.md`27- `reports/AGENT_RUNS.md`28- `reports/INTAKE_GATE.md`2930## Procedure31321. Inventory all provided problem statements, PDFs, DOCX files, spreadsheets, CSV files, images, and downloaded resources.332. Extract the exact contest title, background, top-level questions, attachment descriptions, submission requirements, and known constraints.343. Treat only explicitly numbered top-level questions as `ques1`, `ques2`, etc. Do not inflate small details into fake subproblems.354. If local RAG is available, query `cumcm_problems` or `mcm_icm_problems` for similar task wording, attachment patterns, and hidden scoring cues. Record only sourced hits using `../_references/rag_usage_contract.md` and `../_references/source_quality_policy.md`; core problem facts should come from the current files or `S` official sources, not unsourced memory.365. Audit each data file for rows, columns, field meaning, units, missing values, abnormal values, duplicate records, encodings, and whether it can support each subproblem.376. Record ambiguities and risks in `WORKFLOW_STATE.md`; mark any unreadable or semantically unclear data as `HIGH` risk.387. If using Codex subagents, run `problem-analyst` and `data-auditor` independently, then summarize their outputs into `reports/AGENT_RUNS.md`.3940## PROBLEM_BRIEF.md Structure4142```markdown43# Problem Brief4445## Contest And Task46## Source Files47## Top-Level Questions48## Inputs And Outputs49## Constraints And Evaluation50## Dependencies Between Questions51## Ambiguities And Risks52## Next Modeling Needs53## Local RAG Evidence Used54```5556## DATA_AUDIT.md Structure5758```markdown59# Data Audit6061## File Inventory62## Table And Field Summary63## Missing / Abnormal / Duplicate Data64## Unit And Encoding Issues65## Usable Variables By Question66## Derived Variables Needed67## Blocking Risks68```6970## Gate7172Write `reports/INTAKE_GATE.md` with `PASS`, `CONDITIONAL_PASS`, or `FAIL`.7374Do not proceed to modeling on `FAIL`. On `CONDITIONAL_PASS`, list exactly what the model strategy stage must handle.
zklzzklzkl