Pre-Submission Triage Skill

Purpose

The pre-submission gate scripts/check_pre_submission.py mechanically emits PASS/WARN/FAIL, but cannot judge which WARNs are acceptable and which must not be ignored. This skill cross-references gate output against known patterns, classifies each finding as "must-fix / needs-review / acceptable", presents a judgment summary to the author, and records approved judgments with rationale in a log. The goal is to prevent ad-hoc, person-dependent WARN handling and overlooked judgments.

Trigger Conditions

• When check_pre_submission.py is run before submission or manuscript build
• When the author asks how to handle WARN/FAIL output from the gate
• When reviewing results of chapter-completion or periodic quality checks

Principles

• FAIL always blocks: never classify a FAIL as "acceptable"; fix and re-run the gate
• "Acceptable" WARNs require rationale: a WARN that matches no known pattern below falls to "needs-review"
• Record only after author approval: present the classification and obtain approval before logging. The agent must never silently dismiss a WARN on its own

Workflow

[Step 1: Run gate] -> [Step 2: Match known patterns] -> [Step 3: Present judgment summary] -> [Step 4: Author approval -> Log]

Step 1: Run the Gate

python3 scripts/check_pre_submission.py <paper-id> --repo-root <paper-repo-root>

If every check is PASS with no WARN, report "gate passed" and stop. Otherwise proceed to Step 2.

Step 2: Match Findings Against Known Patterns

literature-grounding

Output pattern	Classification	Criterion
no matching literature/papers/*.md artifact (FAIL)	must-fix	No artifact for the citation; create papers/*.md
status=manual-stub	acceptable	Legitimate grounding form for books/theses where PDF auto-fetch is impossible (fact-grounding-rule §2-B-3). Condition: the papers/*.md records page-anchored excerpts (page-verified excerpts); without excerpts, treat as needs-review
status=abstract-only	needs-review	Visually confirm whether the manuscript claim depends only on abstract-level content; if dependence is strong, upgrade to full-text or manual-stub with page excerpts
status=full-text but no PDF at _downloads/	must-fix	The Markdown note is a transcription, not the original. Place the PDF at _downloads/{slug}.pdf and re-audit. If there is no redistribution right, drop to manual-stub
bot-challenge … human handoff required (download script)	must-fix (human task)	Automated fetch blocked by reCAPTCHA/Cloudflare. Wait until the user saves the PDF to _downloads/{slug}.pdf via browser. Do not retry the same URL or hunt unofficial mirrors (pdf-paper-ingestion Step 1b)

fact-grounding

Output pattern	Classification	Criterion
Paragraphs containing only headings, section numbers, version numbers, or publication years	acceptable	Numeric-pattern false positive
Count mentions in qualitative text (e.g., "four mechanisms")	acceptable	Design constants, not quantitative claims
Quantitative claims (statistics, ratios, multipliers) with no Harvard citation, footnote, docs/data/ reference, or table reference	must-fix	Add an evidence anchor

output-boundary

Output pattern	Classification	Criterion
Internal workflow terms (TDD, test cases, internal file paths) leaking into chapters	must-fix	Never ship internal process vocabulary in the public manuscript; do not treat as WARN

citation-format

Output pattern	Classification	Criterion
In-text citation not registered in the matrix	must-fix	Register in literature-matrix.md and ground via papers/*.md
Undefined footnotes / orphan references	must-fix	Restore 1-to-1 consistency between text and references

Step 3: Present the Judgment Summary

Present in chat using this format:

## Triage Result (<paper-id>, YYYY-MM-DD)

- Gate totals: PASS=n / WARN=n / FAIL=n
- Classification: must-fix n / needs-review n / acceptable n

| Check | Target | Classification | Rationale | Action |
| :--- | :--- | :--- | :--- | :--- |
| literature-grounding | Wood et al. (1976) | acceptable | manual-stub + page-verified excerpts at p.90/98 | none |
| fact-grounding | ch3 §3.2 heading | acceptable | section-number false positive | none |
| fact-grounding | ch5 paragraph X | must-fix | quantitative claim without anchor | add citation footnote |

Step 4: Author Approval and Logging

1. Present the summary and obtain the author's approval for the classifications (especially "acceptable" and "needs-review")
2. After approval, append the judgments to docs/<paper-id>/design/logs/pre-submission-triage-log.md:

## YYYY-MM-DD — pre-submission triage (<paper-id>)

- Gate: PASS=n WARN=n FAIL=n / Classification: must-fix n, needs-review n, acceptable n

| Check | Target | Classification | Rationale | Action |
| :--- | :--- | :--- | :--- | :--- |

3. After completing must-fix items, re-run the gate and append the result to the same log

Outputs

• Triage judgment summary (presented in chat)
• docs/<paper-id>/design/logs/pre-submission-triage-log.md (audit trail of judgments)