论文评审与标注
Generate an evidence-backed, concise thesis audit report. Treat original Word files as protected artifacts: only use the supplied evidence bundle and preview tools.
Required Safety Boundary
- Do not request, open, parse, save, or edit the original
.docor.docxfile. - Use only a
document_id, page-preview tools, and read-only evidence returned by the protected document service. - Never declare a format issue confirmed from a page image alone or from a document property alone.
- Never insert issue IDs, highlight text, or rewrite any content in the student thesis.
- If the visual and structural evidence conflict, report
需复核; do not guess and do not create a Word or PDF annotation.
Workflow
- Confirm that the evidence bundle includes a source SHA-256, template ID, page previews, structural facts, and visual observations.
- Read evidence-bundle-schema.md, template-annotation-requirements.md, rule-applicability-and-confidence.md, and the applicable rule pack. For the initial template, use
rules/software-development-thesis-template.json. - Inspect full-page previews first. Request high-DPI crops for any candidate issue involving punctuation, spelling, caption placement, page number, or visual overflow.
- Classify each conditional rule as
applicable,not_applicable, oruncertainbefore checking it. Record the reason and do not treatnot_applicableas a fault. - Record each visual result as
passorissue, with page, locator, category, confidence, short evidence, and crop reference. Do not infer text from low-resolution images. - Obtain protected-service
annotation_resolutions. Usetext_anchoronly after a unique original-text anchor is validated; usesection_pageonly after the section page is validated; otherwise usereport_only. - Run
scripts/audit_evidence.pywith the evidence bundle, rule pack, and output directory. - Deliver exactly three user-facing files:
学生论文_Word批注版.docx,学生论文_标注审查版.pdf, and论文格式审查评价表.md.
Decision Policy
| Visual evidence | Structural evidence | Report status | Action |
|---|---|---|---|
| Issue | Issue | 已确认 |
Eligible for safe annotation only after locator validation |
| Missing, unclear, or low confidence | Issue | 疑似问题 |
Request a crop or human review |
| Issue | Pass | 需复核 |
Report the conflict; do not call it an error |
| Pass | Issue | 需复核 |
Report the conflict; do not call it an error |
| Issue | No matching property | 视觉疑点 |
Report layout concern only |
Use 已确认 only when the rule is applicable, both channels mark the same locator and category as an issue, and both confidence values are at least 0.90. A missing required section uses the table-of-contents or section-list observation as its visual evidence. not_applicable rules never enter the issue table.
Evaluation Table Requirements
The audit script is the sole report formatter. Its output must be named exactly 论文格式审查评价表.md and contain:
- A deterministic report fingerprint, source SHA-256, template ID, and rule version. Do not include wall-clock time or machine-specific paths.
- Total issue counts split into exact annotations, section-page prompts, and report-only items.
- A compact issue index at the beginning with page and Markdown line number.
- Detailed report-only issues first, then section-page prompts, then exact annotations.
Keep 问题 and 指导建议 concrete and short. For a report-only issue, state why it was not annotated. For a section-page prompt, state clearly that it does not identify a specific sentence or object.
Resources
scripts/audit_evidence.py: deterministically evaluate a prepared evidence bundle and write the Markdown evaluation table.scripts/run_stability_test.py: run the fixture three times and fail unless the reports are byte-identical.references/evidence-bundle-schema.md: protected-service input contract.references/software-development-thesis-template.md: initial template rules and evidence expectations.references/template-annotation-requirements.md: requirements extracted from all 36 Word comments in the supplied template.references/rule-applicability-and-confidence.md: conditional-rule, confidence, error-rate, and scoring policy.references/annotation-safety-contract.md: mandatory Word/PDF annotation and locator-validation contract.rules/software-development-thesis-template.json: machine-readable initial rule pack derived from the supplied template.assets/terminology-template.csv: project-specific terminology input template.
Commands
py -3.14 scripts\audit_evidence.py `
--evidence tests\fixtures\evidence-bundle.json `
--rules tests\fixtures\rules.json `
--output .\out
py -3.14 scripts\run_stability_test.py
The second command must pass before treating any report format change as complete.
