본문으로 건너뛰기
cccskills
무료GitHub에 공개

submission-packager

Package a G6-approved mathematical-modeling paper and its exact supporting code, data, results, and figures without changing their behavior or the workspace.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

설치 방법 보기

포함된 파일(2)

  • SKILL.md4.3 KB
  • scripts/package_submission.py26.6 KB

SKILL.md(원문)

설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.

Purpose

Create one paper Markdown and one 支撑材料/ directory after the submission audits. Preserve selected code, inputs, and outputs under their workspace-relative paths. The packaged code must retain its module imports, CSV parsing, and normal working-directory paths. A packaging manifest stays outside the package at planning/submission_packaging_manifest.json.

Preconditions

  • planning/session_config.json selects rigor_profile: submission.
  • Every planning/manifests/Qx.json is at G6 and allows final assembly.
  • paper/audits/cross_media_consistency_audit.md, paper/audits/completeness_audit.md, and paper/qa_report.md each contain an explicit Verdict: PASSED line. Ask the relevant auditor to record the verdict when an older report omits it; do not infer passage from file existence.
  • The paper and freezes have not changed after those audits. Renew affected audits when they have.
  • Each Qx has a current frozen_numbers.json. Its source files identify one successful experiment round; that round's run_summary.json lists scripts, inputs, and outputs as workspace-relative paths. Update canonical run evidence through the normal workflow when a field is missing. Never select a round by its number alone.
  • A final paper/main.md exists, or supply another Markdown with --paper.

If the human explicitly authorizes packaging before G6, record their choice and rationale as a DECIDED human packaging_waiver in a decision ledger. Pass its decision_id with --allow-unaudited. The package manifest records this waiver and every failed gate or audit; never label it G6-approved.

Workflow

  1. Run python <skill-dir>/scripts/package_submission.py --workspace <ws> with Python 3.9 or newer. This is a dry run and prints every copied source, its destination, blocked items, and a SHA-256 plan digest. Check the actual contest's submission format and present the full inclusion list to the human.
  2. After the human confirms that exact list, run the same command with --confirm <plan digest>. A changed source or plan blocks packaging. --out selects another output directory. --force replaces only a previous, unchanged package recorded in the external manifest.
  3. The script copies selected source files into a staging directory, preserves their workspace-relative paths, rewrites only the packaged Markdown's image paths, checks Python syntax, then moves the verified stage into place. It rejects output paths containing the workspace or selected sources. The original workspace and the frozen sources stay untouched.
  4. When a safe reproduction command is available, pass --verify-command 'python code/Q1/run_all.py' (repeat per required command). Commands execute from the packaged 支撑材料/ directory after copied frozen outputs are removed, so they must recreate those outputs. Numerical JSON claims with a $.path locator are checked against frozen values using --atol and --rtol. Review command side effects before running. A missing command or unsupported locator is reported as UNVERIFIED, never as a runtime pass. MATLAB execution requires a compatible local runtime and an explicit command.
  5. Review the external manifest, package file hashes, runtime status, and actual Markdown image links. Only describe the package as runnable when the required reproduction checks passed.

Selection and limits

  • Code comes from run_summary.json and its local Python/MATLAB dependencies. Inputs come from the summary and literal data paths in selected code. Outputs come from the summary and frozen sources. Paper images come from references in the delivered Markdown.
  • Preserve filenames and directories. Do not consolidate modules, prune comments, convert CSV files, rename data, or copy exploratory rounds, decision ledgers, audit files, or logs into 支撑材料/.
  • The paper is the only Markdown at the package root. Local paper images are copied beneath 支撑材料/ and their references in the packaged paper are updated. Unresolved image paths block packaging.
  • The manifest records evidence hashes, exact source-to-destination mapping, gate and audit evidence, selected frozen rounds, the confirmation digest, package hashes, and runtime status.

리뷰

아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.

댓글

로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.

아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.

보내면 바로 공개됩니다. 이메일 주소나 전화번호 같은 개인 정보는 쓰지 마세요. URL은 2개까지입니다.

0 / 1,000자

이 양식은 reCAPTCHA로 보호되며 Google 개인정보처리방침 및 서비스 약관이 적용됩니다.

같은 저장소의 스킬

Detect whether approved modeling code is Python or MATLAB/Beita Tianyuan and route it to the matching reviewer using the compact named-check review contract.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

zhnnky329/MathModeling-skills1.1천2026년 9월 24일 업데이트

Audit whether the semantic evidence required by the active lean or submission profile exists and is current, without requiring one verbose artifact per skill or an arbitrary number of pass bullets.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

zhnnky329/MathModeling-skills1.1천2026년 9월 24일 업데이트

Run scoped or final cross-media consistency checks for mathematical-modeling artifacts, comparing canonical numbers, symbols, parameters, decisions, files, and paper claims without performing full-workspace audits for low-risk changes.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

zhnnky329/MathModeling-skills1.1천2026년 9월 24일 업데이트

Map contest attachments to subquestions, audit and clean raw data, and emit one reusable data profile with quality, coverage, imbalance, concentration, and method-readiness evidence for downstream risk screening.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

zhnnky329/MathModeling-skills1.1천2026년 9월 24일 업데이트

Build one compact choice card at a genuine mathematical-modeling judgment point. Use before method screening, after a meaningful experiment, or before final claim/freeze approval so the human chooses the trade-off while AI handles mechanical consequences.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

zhnnky329/MathModeling-skills1.1천2026년 9월 24일 업데이트

Plan the smallest set of diagnostic, comparison, paper, and appendix figures or tables needed to support verified mathematical-modeling decisions and claims.

한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.

zhnnky329/MathModeling-skills1.1천2026년 9월 24일 업데이트

zhnnky329의 스킬 모두 보기

이 스킬의 문제 신고