sopaper-evidence（纸质证据）

Sopaper Evidence

Sopaper Evidence is an evidence-first research skill. Its job is to build a reliable evidence pack before supporting any downstream paper outline, abstract, related work summary, experiment plan, or draft section.

Version: v1.0.0

Upstream source

Canonical repository: https://github.com/sheepxux/SoPaper-Evidence

This published skill bundle includes the helper scripts it references under scripts/. The GitHub repository remains the public source of truth for releases, examples, and issue tracking.

Use this skill when

• The user wants to turn a project into a paper without inventing evidence
• The task requires finding prior papers, datasets, benchmarks, baselines, or case studies
• The task requires mapping claims to verified sources
• The task requires identifying evidence gaps before writing
• The user wants related work or experiment planning grounded in real sources

Hard rules

• Do not fabricate papers, authors, venues, dates, citations, datasets, benchmarks, experiments, or numerical results
• Prefer primary sources over summaries, reposts, or blog interpretations
• Separate verified facts from inference and open questions
• If evidence is missing, say it is missing and recommend what to collect next
• Do not state that the user's method outperforms baselines unless there is explicit evidence
• Every writing-oriented output must be traceable to evidence items

Source priority

Use the highest-quality source available for each claim.

1. User-provided project artifacts: experiment logs, tables, code, configs, internal notes
2. Primary external sources: papers, official docs, benchmark leaderboards, dataset pages, project repos
3. Secondary summaries: blogs, news posts, third-party explainers

Read references/source-priority.md when source quality or conflicts matter. Read references/input-schemas.md when stronger input structure is needed before running the workflow.

Core workflow

1. Scope the task

Collect or infer:

• Project name
• Research topic
• Core problem
• Method summary
• Existing evidence and file paths
• Target venue or paper style if known

If the project scope is unclear, produce a short working scope and label assumptions.

2. Search for evidence

Search for:

• Prior work
• Benchmarks and datasets
• Baseline methods
• Comparable case studies
• Official metrics definitions
• Relevant project artifacts in the local repository

For each source, capture the title, URL or path, source type, and why it matters.

Use references/prior-work-search-playbook.md for a repeatable search process. For OpenClaw-specific work, use references/openclaw-evidence-playbook.md.

3. Verify and classify

For each evidence item, classify it as:

• verified_fact
• project_evidence
• inference
• unverified

Do not merge these labels. If a statement depends on inference, say so explicitly.

4. Extract structured evidence

Use the schema in references/evidence-schema.md.

At minimum, extract:

• Claim or observation
• Source
• Evidence type
• Scope and limitations
• Relevance to the user's paper

5. Build the evidence map

Organize findings into:

• related_work
• datasets_and_benchmarks
• baselines
• case_studies
• project_results
• claim_to_evidence
• evidence_gaps

Use assets/claim-evidence-map-template.md when the user needs a reusable deliverable. Use assets/related-work-matrix-template.md when comparing papers, baselines, and benchmark coverage. Use assets/experiment-gap-report-template.md when the task requires prioritizing missing experiments before drafting. Use bundled scripts/build_evidence_ledger.py when the user already has markdown notes or source lists and needs a first-pass evidence ledger. Use bundled scripts/generate_search_plan.py when the user starts only with a topic and needs a first-pass evidence search plan. Use bundled scripts/generate_topic_claims.py when the user starts only with a topic and needs a cautious structured claims draft. Use bundled scripts/search_external_sources.py when the user needs a first-pass source list from a topic or search plan. Use bundled scripts/fetch_external_sources.py when raw URLs should be converted into structured source-note drafts before review. Use bundled scripts/verify_source_notes.py when fetched notes should be conservatively upgraded into page-level verified facts or reviewed primary-source summaries before entering the ledger. Use bundled scripts/run_evidence_pipeline.py when the user already has source files, claims, and optional result artifacts and wants one end-to-end draft pack. Result artifacts may be structured markdown, .csv, .tsv, or .json, and multiple result artifacts can be fused into aggregate project evidence. Use bundled scripts/bootstrap_claim_map.py when the user already has a claims list and a ledger draft and needs a first-pass claim map. Use bundled scripts/triage_evidence_gaps.py when the user needs a first-pass blocker/major/minor gap report from the current claims and evidence ledger. Use bundled scripts/review_comparison_fairness.py when the user needs a dedicated fairness check on comparative claims, baseline breadth, metric grounding, and scope alignment. Use bundled scripts/run_topic_evidence_pipeline.py when the user wants the full topic-driven workflow from theme to search plan, source list, fetched notes, ledger, claim map, and gap report. Use bundled scripts/validate_input_bundle.py when the user has partially structured inputs and needs a quick schema check before running the pipeline.

6. Support writing

Only after the evidence map is complete, support tasks such as:

• contribution candidates
• related work summary
• abstract support points
• experiment plan
• paper outline

Before writing, run the checks in references/claim-audit-rules.md. Use assets/paper-outline-from-evidence-template.md when the user needs a draft-safe paper structure.

Output requirements

Unless the user asks for something else, default to this output shape:

1. Evidence brief
2. Key sources
3. Claim-to-evidence map
4. Evidence gaps
5. Safe writing notes
6. Experiment gap report when blocker gaps exist

See the example set in:

• examples/openclaw-input.md
• examples/openclaw-search-plan.md
• examples/openclaw-evidence-brief.md
• examples/openclaw-claim-map.md
• examples/openclaw-gap-report.md
• examples/openclaw-source-note.md
• examples/openclaw-source-list.md
• examples/openclaw-ledger-draft.md
• examples/openclaw-claims.md
• examples/openclaw-claims-structured.md
• examples/openclaw-claim-map-draft.md
• examples/openclaw-gap-report-draft.md
• examples/openclaw-paper-outline.md

Writing constraints

When supporting downstream paper writing:

• Tie each major claim to one or more evidence items
• Avoid precise quantitative wording unless the number is verified
• Mark missing comparisons, missing ablations, and missing real-world validation
• Prefer conservative wording over overstated conclusions

OpenClaw-specific guidance

When the user is working on OpenClaw or a similar embodied AI / robotics project, prioritize:

• manipulation benchmarks
• long-horizon task evidence
• policy or planner comparisons
• real-world versus simulation evidence
• ablations on perception, planning, or control components

Do not assume OpenClaw has capabilities, datasets, or benchmark wins unless they are present in project artifacts or verified sources. Use references/benchmark-baseline-checklist.md before accepting benchmark-fit or baseline coverage claims. Use references/evidence-gap-triage.md when deciding whether to keep drafting or stop and report blockers.