Nsfc Proposal

Inputs Required

Collect before execution:

• Project basics: title, discipline code, project type, research attribute, duration, budget.
• Existing materials: draft files, prior work, platform/conditions, related projects.
• User constraints: word targets per section, preferred P2 sub-structure, H/O/RC/KSQ mapping count.

Tooling Rules

Academic literature retrieval must use paper-search MCP tools:

• paper-search_search_pubmed (primary)
• paper-search_search_semantic (supplement)
• paper-search_search_google_scholar (coverage)
• paper-search_search_biorxiv / paper-search_search_medrxiv (recent preprints)

Runtime name mapping (for environments using MCP namespaced tools):

• paper-search_search_pubmed -> mcp__paper-search__search_pubmed
• paper-search_search_semantic -> mcp__paper-search__search_semantic
• paper-search_search_google_scholar -> mcp__paper-search__search_google_scholar
• paper-search_search_biorxiv -> mcp__paper-search__search_biorxiv
• paper-search_search_medrxiv -> mcp__paper-search__search_medrxiv

Do not use generic web search/fetch tools for citation evidence in proposal claims.

Non-Conflict Canon (Authoritative)

Apply these resolutions when references conflict:

1. No-bullet narrative applies to proposal body sections only; diagnostics/review reports may use structured lists.
2. Interaction extras (reverse questioning, suggested follow-up questions, extended thinking) are optional by context, not mandatory on every response.
3. Merge order is fixed: references at the end of final merged manuscript.
4. V-01 consistency validation must not require nonexistent mapped_to_h fields on SQ nodes. Validate by traversing H and KSQ links from SQ references.
5. P2 should not include numbered literature markers; citation numbering is restricted to P1.

Execution Workflow

Write Mode

Follow phased gates in order:

1. Phase 0: initialize project profile, section targets, mapping cardinality.
2. Phase 1: write P1 with full citation pipeline and verification.
3. Phase 2: write P2 and enforce strict H/O/RC/KSQ one-to-one mapping.
4. Phase 3: write P3 with feasibility evidence linked to methods.
5. Phase 4: write P4 and compliance declarations.
6. Phase 5: budget narratives and traceability to methods.
7. Phase 6: Chinese/English abstracts generated last from full content.
8. Phase 7: global consistency + review + merge.

At each phase:

• snapshot
• sync required state files
• halt for user confirmation

Polish Mode

1. Import draft and split into atomic section files by original heading hierarchy.
2. Generate strict review report first (polish_review_report).
3. Agree priority with user (rewrite vs polish vs keep).
4. Revise section by section following issue order:
   • academic design/hypothesis
   • consistency
   • writing style
   • format/length
5. Run global consistency repair and full review.
6. Merge final output.

State and Artifacts

Maintain and sync after each section edit:

• data/consistency_map.json
• data/literature_index.json
• context_memory.md
• project_state.json
• history_log.json

Any missing sync blocks phase progression.

• Use dual-track verification for citations: provide MCP retrieval cache in data/mcp_literature_cache.json and run online validation without --offline whenever network is available. Final gate should enforce --require-mcp.

Quality Gates

Block progression when any of the following fails:

• ERROR-level consistency rules.
• Unverified references in P1 citation set.
• Citation-index-reference matrix mismatch.
• Any D-grade in global review dimensions.
• More than 3 C-grade dimensions in global review.
• Page estimate beyond configured hard limit.

Use atomic gate command for final checks:

• python scripts/state_manager.py --root . gate-check --sections-dir sections --index data/literature_index.json --p1 sections/P1_立项依据.md --ref sections/REF_参考文献.md --mcp-cache data/mcp_literature_cache.json --mcp-ttl-days 30 --require-mcp

Failure handling playbook:

• failed_at=sync: run sync-all --auto-fix, then re-run gate-check.
• failed_at=citation: repair index/cache, re-run verify-all --require-mcp, then gate-check.
• failed_at=matrix: run matrix-check and reorder, then gate-check.
• failed_at=review: fix D/C dimensions from review report, then gate-check.

References

Load only what is needed:

• references/00_设计方案_总览.md
• references/01_目录结构与配置.md
• references/02_核心机制.md
• references/03_写作规范与反AI.md
• references/04_文献管理.md
• references/05_Write_Mode流程.md
• references/06_Polish_Mode流程.md
• references/07_自审与评审模块.md
• references/08_脚本清单与合并规则.md
• references/09_交互规范与回复模板.md

Output Contract

Deliverables should include:

• section files under sections/
• updated state and data files
• review reports in data/
• merged manuscript in output/ (md/docx if requested)

When reporting to user, state:

• what was changed
• which gate passed/failed
• what is blocked and exact unblock action

Script Entry Points

Use scripts under scripts/ from proposal project root:

• python scripts/state_manager.py --root . init
• python scripts/state_manager.py --root . load --section P1 --minimal
• python scripts/state_manager.py --root . write-cycle --section P1 --token-budget 4000
• python scripts/state_manager.py --root . sync-all --auto-fix
• python scripts/state_manager.py --root . gate-check --sections-dir sections --index data/literature_index.json --p1 sections/P1_立项依据.md --ref sections/REF_参考文献.md --mcp-cache data/mcp_literature_cache.json --mcp-ttl-days 30 --require-mcp
• python scripts/consistency_mapper.py --path data/consistency_map.json validate
• python scripts/consistency_mapper.py --path data/consistency_map.json validate-one V-01
• python scripts/citation_validator.py verify-all --index data/literature_index.json --p1 sections/P1_立项依据.md --mcp-cache data/mcp_literature_cache.json --mcp-ttl-days 30 --require-mcp --manual-review data/manual_review_queue.json --log data/verification_run_log.json
• python scripts/citation_validator.py verify-entry --index data/literature_index.json --p1 sections/P1_立项依据.md --ref-number 1 --mcp-cache data/mcp_literature_cache.json --require-mcp
• python scripts/citation_validator.py matrix-check --p1 sections/P1_立项依据.md --index data/literature_index.json --ref sections/REF_参考文献.md
• python scripts/humanizer_zh.py scan-all sections
• python scripts/diagnosis_engine.py full-review --sections-dir sections --consistency data/consistency_map.json --index data/literature_index.json --p1 sections/P1_立项依据.md --ref sections/REF_参考文献.md --output data/diagnosis_report.json
• python scripts/diagnosis_engine.py polish-review --sections-dir sections --consistency data/consistency_map.json --index data/literature_index.json --p1 sections/P1_立项依据.md --ref sections/REF_参考文献.md --json-output data/diagnosis_report.json --md-output data/polish_review_report.md
• python scripts/section_merger.py validate-order --sections-dir sections
• python scripts/section_merger.py merge --sections-dir sections --output output/申请书_合并.md
• python scripts/section_merger.py merge --sections-dir sections --only P1_立项依据.md,P2,REF_参考文献.md --output output/阶段稿_合并.md
• python scripts/word_counter.py summary sections

These scripts are production-ready workflow utilities for iterative proposal drafting and polishing.

Regression Tests

• python3 -m unittest discover -s tests -p 'test_*.py'