Engineering Validation

Use this skill before saying a paper edit is complete, fixed, or ready.

Core Stance

• No verification, no completion claim.
• A dry read is not a build check.
• Report exactly what was run and what was not run.
• Treat evidence anchors, citations, figures, tables, and PDF output as part of manuscript correctness.
• For existing paper repos, read the current live draft and project-specific instructions before validating.
• This is readiness validation, not truth certification. Never certify experiment truth, citation truth, or venue compliance without the relevant source files, tools, and checks.

Boundaries

• Use this skill for readiness checks, not for drafting new manuscript prose.
• Use engineering-writing when the user needs new section structure or claim-evidence planning before validation.
• Use engineering-polishing for prose-only improvements.
• Use engineering-figure-table for visual design before final consistency checks.
• Use engineering-response for reviewer/advisor reply drafting before final response package validation.

Status Vocabulary

• PASS: checked directly and no issue found.
• FAIL: checked directly and an issue was found.
• PARTIAL: checked only in part; name the unchecked part.
• NOT_RUN: check was relevant but not run.
• UNKNOWN: cannot determine from provided material.

Overall readiness can only be READY, NOT_READY, or CANNOT_DETERMINE.

READY is allowed only when every required check for the user's stated goal is PASS. If any required check is FAIL, PARTIAL, NOT_RUN, or UNKNOWN, overall readiness must be NOT_READY or CANNOT_DETERMINE.

LaTeX Consistency Checker

Run python3 scripts/paper_check.py path/to/main.tex [--bib refs.bib] [--json] (the script lives in this skill's scripts/ directory; standard library only). It follows \input/\include and reports, with file:line:

• errors: duplicate \label keys, references to undefined labels, citation keys missing from the bibliography, a declared bibliography that cannot be read;
• warnings: floats never referenced, figures or tables first referenced out of numeric order (in source, and in the compiled numbering when a .aux file is found or passed with --aux), success rates that are impossible for the stated number of trials (heuristic; it skips aggregated values and reports ambiguous trial counts as info), average or total columns that do not match their row, and prose values that contradict the table cell they name (heuristic; silent when the match is ambiguous);
• info: uncited bibliography entries, prose percentages that appear in no table, and anything it could not check (an input file outside the project, a dynamic macro argument).

It skips macro definitions, verbatim-like environments, and \iffalse blocks, and treats only \% and siunitx forms as percentages. Without a .aux file it reads source order only. It does not expand general macros, so numbers defined through \newcommand are not checked. Report its output verbatim as one check in the readiness table; a clean run is evidence for these checks only.

When to Open Extra Files

Workflow

1. State the validation goal being interpreted from the user request.
2. Identify manuscript root and live draft.
3. Read local instructions and relevant build files.
4. Check structure anchors: thesis, contribution list, main results, ablation, limitations.
5. Check story spine: Abstract and Conclusion claims must have Methods, Experiments, figure/table, and boundary anchors.
6. Check figures, tables, captions, labels, and category/metric names. Check that floats are first referenced in numeric order, that no \label is duplicated, and that every \ref resolves to the intended float; wrong or out-of-order table references were flagged by real reviewers.
7. Run the smallest relevant build and consistency commands.
8. If tools or files are unavailable, mark command-dependent checks NOT_RUN and perform only the file-based or local-text audit that evidence allows.
9. For venue readiness, report whether official policy was checked, including source URL and source date. If not checked, venue compliance is NOT_RUN or PARTIAL. Static venue profiles count only as STATIC_SUMMARY_NOT_LIVE_CHECK, not as official policy verification.
10. Report evidence, failures, and unverified items.

Default Output

Validation result
- Goal interpreted as:
- Build: PASS / FAIL / NOT_RUN / UNKNOWN
- References: PASS / FAIL / PARTIAL / NOT_RUN / UNKNOWN
- Figures/tables: PASS / FAIL / PARTIAL / NOT_RUN / UNKNOWN
- Claim-evidence audit: PASS / FAIL / PARTIAL / NOT_RUN / UNKNOWN
- Story-spine audit: PASS / FAIL / PARTIAL / NOT_RUN / UNKNOWN
- AI/disclosure check: PASS / FAIL / PARTIAL / NOT_RUN / UNKNOWN
- Venue readiness: PASS / FAIL / PARTIAL / NOT_RUN / UNKNOWN
- Private/stale wording scan: PASS / FAIL / NOT_RUN / UNKNOWN
- Overall readiness: READY / NOT_READY / CANNOT_DETERMINE

Evidence
- Commands run:
- Files inspected:
- Outputs observed:
- Failures:
- Not verified:
- Official policy checked:
- Official policy source/date:
- Official page opened this run:
- Policy excerpt checked:
- Profile-only status:

Blocking issues
- ...

Next actions
- ...