审计报告：research-units-pipeline-skills
总体结论
- Verdict:
WARN - Evidence level:
L3 - Core L3 completed: yes (4/4 tasks)
- Website recommendation: 可推荐，需标注限制条件
- Reason: 架构设计优秀（checkpoint gating, provenance tracking, multi-pipeline design），4 个核心模块均有 L3 证据。但网络依赖和 prompt 质量依赖是两个实际风险，需要用户在部署前了解。
证据来源
- Evidence tasks:
evidence/tasks.yaml - Evidence scorecard:
evidence/scorecard.md - Evidence index:
evidence/evidence-index.md - Case run log:
run-log.md - Raw outputs:
outputs/
通用门槛
| 门槛 | 结果 | 证据 |
|---|---|---|
| Public source | PASS | GitHub: WILLOSCAR/research-units-pipeline-skills (MIT) |
| Version pinned | PASS | candidate.yaml — git main branch |
| Install path clear | PASS | Prompt-driven; no binary installation required |
| Permissions transparent | PASS | candidate.yaml — network: true, reads/writes files, runs scripts |
| Real research task | PASS | 4 tasks across Chemistry/Physics/Medicine/Biology |
| Not pure marketing shell | PASS | evidence/evidence-index.md — all tasks have input/output evidence |
模块结论
| 模块 | 结论 | 核心主张 | 最高深度 | 证据 | 风险 |
|---|---|---|---|---|---|
| literature-discovery | WARN | Multi-route literature discovery with provenance | L3 | lit-disc-01 — offline import works, online unverified | Network-dependent online routes |
| paper-reading-kb | PASS | Structured paper notes with method/results/limitations | L3 | paper-kb-01 — complete notes with 5 bullets, method, 4 results, 3 limitations | Prompt quality affects limitation specificity |
| research-writing | PASS | Evidence-bounded writing with citation scope | L3 | writing-01 — 293-word subsection, caveat-aware | Numbers approximate without full trial text |
| project-management | PASS | Auditable pipeline execution with checkpoints | L3 | pm-01 — workspace initialized, 6-unit pipeline, C2 gating | Full pipeline execution not tested end-to-end |
输出质量
| 输出 | 质量标签 | 证据 | 修改负担 |
|---|---|---|---|
| papers_raw.jsonl | WARN | lit-disc-01 — 10 records with DOI + provenance | Low; needs online retrieval for >=1200 |
| paper_notes.jsonl | PASS | paper-kb-01 — structured, specific, non-boilerplate | Low; fulltext mode would enrich results |
| sections/S3.1.md | PASS | writing-01 — publishable draft quality | Low; update numbers with full trial data |
| workspace artifacts | PASS | pm-01 — complete initialization | Low; subsequent units need network |
安全说明
- Network: Required for online literature retrieval (arXiv API, Semantic Scholar). Offline import mode available as fallback.
- Local files: Reads/writes workspace files within defined directory structure. No path traversal risk observed.
- Scripts: Python scripts for scaffolding/validation only; not the primary execution path.
- API keys: None required for core functionality.
- External services: arXiv API, Semantic Scholar API — both are free, rate-limited public APIs.
- Data exposure: All test papers are publicly published. No user research data involved.
网站收录说明
- Best for: Academic researchers needing structured, auditable literature survey workflows with evidence traceability. Best fit for those comfortable with a multi-step, checkpointed process.
- Not for: Users wanting one-click paper summaries; users who cannot configure network access for literature APIs; users expecting script-level determinism (skills are prompt-driven).
- Install difficulty: Low (git clone, no compilation). Network configuration may be needed in China.
- Safety level: Medium — prompt-driven skills may exhibit model-dependent quality variance; guardrails are strong but enforcement depends on model compliance.
- Reviewer note: Strong architectural design with genuine research value. WARN verdict reflects environmental limitations (network) and model-dependency risk, not design flaws. Recommend inclusion with caveats.
- Recommendation boundary: 可推荐为"结构化文献综述工具"，不可推荐为"一键论文总结工具"。Detail page 必须说明 checkpoint 机制、网络需求和 prompt-driven 特性。
Summary Constraint: summary.md must not claim stronger recommendations than this review supports.