shield_lock
金大哥
← 返回 Skill 列表
extension
分类: 开发与工程无需 API Key

figuregate

Run local, evidence-grounded publication QA for scientific figures. Use when a user asks to demonstrate FigureGate in WorkBuddy, compile journal figure guidelines, inspect figure resolution/format/physical width, enable conservative visual QA, or apply presentation-safe repairs and recheck the result.

person作者: asdsdfghhubgithub
open_in_new仓库download下载资源包

FigureGate

FigureGate is a local publication gate for scientific figures.

Core rule: LLM understands. FigureGate verifies. Agent fixes.

Vision rule: Vision detects. VLM interprets. Validator decides.

Locate the project

Treat the directory containing this SKILL.md as the project root. Run all commands from that directory. Do not assume a fixed drive letter or user profile.

WorkBuddy screenshot demo

When the user asks for a demo, assignment screenshot, or proof that the Skill works:

  1. Verify .venv\Scripts\python.exe exists. If it does not, follow the installation commands in README.md; do not claim the demo ran.

  2. Run:

    .\demo\run_workbuddy_demo.ps1

  3. Require the script to finish with FIGUREGATE WORKBUDDY DEMO: PASS.

  4. Report the real rule count, before status, after status, and the generated report paths printed by the script.

  5. For a screenshot, keep the final terminal summary visible. Never fabricate a screenshot or a PASS result.

This fast demo uses a checked-in rule-extraction fixture so it can prove the deterministic FAIL -> safe repair -> PASS workflow without model downloads. State that limitation. The accepted real-model demonstrations are separate commands documented in README.md.

Compile real journal guidelines

Use the local OpenVINO backend only when the compatible Qwen model is available:

.\scripts\run.ps1 compile --guideline <guideline.pdf> --backend openvino --model <model-directory> --device CPU --output <rules.json>

Executable rules must retain the guideline page and exact evidence text. Unsafe or unsupported model output stays in validation_issues; never replace it with guessed journal defaults.

Inspect figures

.\scripts\run.ps1 inspect --rules <rules.json> --manifest <figure-manifest.json> --output-dir <output-directory>

Read inspection.json before editing anything. Keep PASS, FAIL, WARNING, NOT_CHECKED, NEEDS_REVIEW, and ERROR distinct.

Enable the optional visual worker only when the isolated vision environment and local models described in README.md exist. A VLM-only negative judgment cannot directly produce FAIL; insufficient or conflicting evidence becomes NEEDS_REVIEW.

Safe repair contract

Automatic edits are limited to explicit presentation-level source parameters supported by evidence, including:

  • figure size;
  • export DPI;
  • output format and filename;
  • explicit font-size parameters;
  • explicit panel labels;
  • clearly recoverable axis visibility.

Never automatically change:

  • data values or numerical results;
  • scientific units by guessing;
  • axis meaning;
  • series semantics;
  • plotting logic or scientific conclusions.

Regenerate from source and run inspection again. Never claim that interpolating a low-resolution raster creates genuine publication resolution. If the source is missing or the safe edit is unclear, return manual guidance instead of editing.

Evidence to return

Return the paths to inspection.json and report.md, summarize the before/after states, and identify the evidence used. If a worker crashes or times out, preserve deterministic results and report the visual subsystem as ERROR.