Pre Submission Report

Pre-Submission Report

Aggregates all quality checks into one dated report. Run before submitting to a journal/conference or sharing with collaborators.

Output Path

Per rules/review-artefact-routing.md (auto-loads in research projects (path-scoped to paper-*/ and paper/)):

• Source slug: pre-submission-report
• Write reports to: reviews/pre-submission-report/YYYY-MM-DD.md inside the project. Path is relative to the research project root, not the Task-Management repo.
• Never at project root (./CRITIC-REPORT.md-style filenames are forbidden — pre-rule layout).
• Idempotency: if today's file exists, append a same-day descriptor ({date}-revision.md, {date}-r2.md, {date}-pre-submission.md) — never overwrite.
• Index update: if reviews/INDEX.md exists, write a one-line entry under "Latest per source" pointing at the new file. Otherwise /review-recap will rebuild the index next time it runs.
• Infrastructure repos (Task-Management, atlas-workspace, etc.): this section does not apply — the path-scoped rule won't load there.

When to Use

• Before submitting a paper to a venue
• Before sharing a draft with supervisors or co-authors
• When the user says "pre-submission check", "is this ready?", "run everything"

Input

• A .tex file path, or auto-detect paper/main.tex in the current project

Critical Rule

Python: Always use uv run python or uv pip install. Never bare python, python3, pip, or pip3. Include this in any sub-agent prompts.

Steps

1. Locate the Paper

If no argument provided, search for the main .tex file:

1. Check paper/main.tex
2. Check paper/*.tex for a file containing \begin{document}
3. Ask the user if ambiguous

2. Integrity Gate (hard gate — must pass before quality checks)

Run these checks first. If any fail, stop and report — do not proceed to quality checks.

1. Placeholder scan — grep the .tex file(s) for TODO, FIXME, XXX, TBD, [INSERT, PLACEHOLDER, Lorem ipsum. Any match is a FAIL.
2. Citation integrity — invoke /bib-validate in verify mode. Every \cite{} key must resolve to a .bib entry. Any missing key is a FAIL.
3. Section completeness — check that all standard sections exist and are non-empty (Abstract, Introduction, and at least one body section before Conclusion/References). An empty or missing section is a FAIL.
4. Broken references — grep for ?? in the compiled PDF output or .log file (undefined \ref{} or \cite{}). Any ?? in output is a FAIL.
5. Anonymity gate (only if the venue is double-blind) — load _shared/double-blind-anonymity-checklist.md and run all P1–P8 paper-side checks. Any FAIL is a hard stop. In particular: P4 (self-citation bib must be blinded if cited paper's author list overlaps the submission's) and P5 (body text must not name authors of self-cited works) — these are the CCS 2026 #1328 desk-reject triggers and require the submission's author list to be loaded from the vault submission frontmatter or prompted from the user. If the artifact has been minted via /anonymous-artifact, also confirm A1–A9 ran clean for that artifact (state file at <project>/.anonymous-artifact-state.json). Skip this entire step only when the user explicitly says "single-blind" or "non-blind".

If any check fails:

INTEGRITY GATE: FAIL

Blockers (must fix before quality checks):
  - [ ] 3 TODO placeholders found (lines 47, 112, 289)
  - [ ] 2 undefined references (\ref{fig:missing}, \cite{nonexistent2024})
  - [ ] Abstract section is empty

Fix these and re-run /pre-submission-report.

If all pass: proceed to Step 3.

3. Run Quality Checks

Two modes:

3a. Sequential (default — fast, deterministic)

Run these in order — each depends on a clean state from the previous:

1. Compilation — invoke /latex on the main .tex file. Record pass/fail and any remaining warnings.
2. Citation audit — invoke /bib-validate --verify-doi (DOI resolution mode catches fabricated entries). Record missing, unused, suspect, and unresolved-DOI keys.
3. Adversarial review — launch paper-critic agent (via Task tool). Capture the CRITIC-REPORT.md score and findings.

3b. Parallel 7-audit fan-out (--parallel flag)

Use when (a) the paper is near submission and you want a comprehensive scan, or (b) the user explicitly asks for the "full pre-submission swarm". Dispatches 7 read-only sub-agents in parallel via the Task tool, then consolidates findings.

Hard rules for parallel mode:

1. All sub-agents are read-only with respect to project files under review — see subagent-write-guard.md rule. They do NOT modify the paper, bib, code, or any other artefact under review; the orchestrator (this skill) decides what to fix. They DO write their own per-agent reports to reviews/<source-slug>/<YYYY-MM-DD-HHMM>.md per each agent's "Log to REVIEW-STATE.md (final step)" instruction — this is the durable record + the INDEX.md stamp that /review-recap reads. The "read-only" scope is the artefact under review, NOT a prohibition on writing the review report itself.
2. Each sub-agent gets the standard forbid-list — no git, no latexmk, no edits to files outside their scope. The forbid-list explicitly carves out the reviews/<source-slug>/ path as a permitted write target (the agent's logging step needs it).
3. Findings consolidate into a P0/P1/P2 fix list before any edits — single triage point, not 13 streams. Sub-agents return structured findings to the orchestrator in addition to writing their report file; the consolidate step uses the structured returns.
4. No edit phase auto-runs — the user reviews the consolidated report and approves which fixes to apply.

The 7 sub-agents:

Always dispatched (13 sub-agents):

Conditional follow-ups (run after parallel fan-out, opt-in):

The conditional follow-ups are NOT in the parallel batch — they're sequential because they may modify files (and so violate the read-only invariant of the parallel sub-agents).

Dispatch protocol:

# Pseudocode for orchestration — see Task tool docs for actual API
parallel_tasks = [
    Agent("bib-verifier", subagent_type="general-purpose",
          prompt=f"Read {paper_path}/references.bib. Run scholarly scholarly-verify-dois on all DOI-bearing entries (batch ≤50). Report unresolvable DOIs, missing DOIs, suspected fabricated entries. READ-ONLY. {forbid_list}"),
    Agent("novelty-scout", ...),
    # ... 5 more
]
# Wait all → consolidate
findings = consolidate_p0_p1_p2(parallel_tasks)

Sub-agents run concurrently — total wall-clock is bounded by the slowest (typically novelty-scout at ~2-3 min via OpenAlex).

Consolidation: the orchestrator merges findings from all 13 sub-agents into a single P0/P1/P2 fix list:

• P0 (block submission): anonymity leaks (#11), fabricated citations (#1, #2), compilation errors (#12), over-page-limit (#12), code-paper mismatches (#8), prose-replication divergence (#9)
• P1 (must fix): unresolved DOIs (#1), claim-verify failures (#2), novelty threats (#3), critic-report Major issues (#4), domain-review math errors (#5), reproducibility issues (#10), referee2-reviewer concerns (#6)
• P2 (should consider): blindspot virtues + minor vices (#7), AI-detect hot zones (#13), critic-report Moderate/Minor issues (#4), novelty positioning (#3)

Code-bearing detection: if the project has non-tex / non-bib files outside paper-*/ and notes/ (typical signal: code/, data/, scripts/, analysis/, src/ directories), enable code-side sub-agents (#8 code-paper-auditor, #9 artifact-coherence-auditor, #10 reproducibility-auditor) and queue /anonymous-artifact as a conditional follow-up.

Edit phase (separate, opt-in): if the user approves any P0/P1 fixes, dispatch a second round of edit-agents with explicit scoped permissions per file — see subagent-write-guard.md for the forbid-list pattern. The orchestrator confirms each edit-agent's scope before dispatch.

Why parallel: the 13 audits are independent — sequential is ~10x slower for the same coverage. Reviewer-pool collision risk is zero (read-only sub-agents).

Skip parallel mode if: paper is in early drafting (use the sequential 3-audit instead — faster feedback for incomplete drafts), or sub-agents fail repeatedly (fall back to sequential).

4. Aggregate Report

Save to log/audits/quality-reports/YYYY-MM-DD_<project-name>.md:

# Pre-Submission Quality Report

**Project:** <project name>
**Date:** YYYY-MM-DD
**File:** <path to main.tex>
**Target:** <venue from project AGENTS.md, or "not specified">

---

## Integrity Gate: PASS / FAIL

- **Placeholders:** 0 found
- **Citation integrity:** all keys resolved
- **Section completeness:** all sections present
- **Broken references:** none

---

## Overall Score: XX/100 — [Verdict]

Verdict uses the quality scoring framework:
- 90-100: Publication-ready
- 80-89: Minor revisions needed
- 70-79: Significant revisions needed
- Below 70: Not ready

---

## Compilation

- **Status:** PASS / FAIL
- **Warnings:** <count>
- **Details:** <brief summary of any issues>

## Citations

- **Missing keys:** <count> — <list>
- **Unused keys:** <count> — <list>
- **Suspect entries:** <count> — <list>

## Adversarial Review

- **Score:** XX/100
- **Key findings:**
  - <finding 1>
  - <finding 2>
  - ...

## Research Quality Score

Load `skills/shared/research-quality-rubric.md` and report the weighted aggregate (X.X / 5.0) with verdict.

## Remaining Issues

| # | Severity | Category | Issue |
|---|----------|----------|-------|
| 1 | High/Medium/Low | Compilation/Citation/Content | <description> |

## Recommendation

**[Submit / Revise / Not ready]**

<1-2 sentence summary of what needs to happen before submission>

5. Present Summary

Display the report path and the summary table to the user. If the recommendation is "Submit", congratulate. If "Revise", list the top 3 issues to fix first.

Error Handling

• If compilation fails after /latex, still run the remaining checks. Mark compilation as FAIL in the report.
• If paper-critic agent fails, note it in the report and base the overall score on compilation + citations only.
• Always produce the report file, even if some checks failed.

Integration

REVIEW-STATE.md propagation (orchestrator-side stamping)

This skill is an orchestrator in the REVIEW-STATE.md schema. As of the 2026-05-19 architecture change, the orchestrator handles all stamping; sub-agents emit directives but do not call the helper themselves.

Required orchestrator behaviour

When constructing prompts for any sub-agent that is a logging tool (paper-critic, referee2-reviewer, peer-reviewer, domain-reviewer, claim-verify, blindspot, fatal-error-check, code-paper-auditor, artifact-coherence-auditor, reproducibility-auditor, code-review), include this line in the sub-agent prompt:

Emit a review-state-stamp directive at the end of your final response per skills/_shared/stamp-directive-spec.md. Set trigger: pre-submission-report (or omit — this orchestrator overrides). Do not call the stamping helper yourself.

After each sub-agent returns

For each agent's return:

1. Write the agent's final response to a temp file (/tmp/pre-submission-<agent>.md).
2. Parse the directive:

   ARGS=$(bash ~/.Codex/skills/_shared/parse-stamp-directive.sh /tmp/pre-submission-<agent>.md)
   

   If parse-stamp-directive.sh exits non-zero, log a warning ("Agent X return did not contain a review-state-stamp directive — INDEX.md not updated for this run") and continue.
3. Verify the .md report file exists; reconstruct from return content if missing:

   VERIFY=$(bash ~/.Codex/skills/_shared/post-dispatch-verify.sh \
       --return-file /tmp/pre-submission-<agent>.md \
       --project "$PROJECT_ROOT" \
       --agent <agent>)
   

   If VERIFY starts with RECONSTRUCTED, append (report reconstructed by orchestrator — agent skipped Write) to the --notes value before stamping. Guards against the blindspot-class failure mode. See log/2026-05-21-blindspot-write-fix.md.
4. Stamp with the orchestrator's --trigger override:

   eval bash ~/.Codex/skills/_shared/review-state-log.sh "$ARGS" \
       --trigger pre-submission-report \
       --source agent \
       --project "$PROJECT_ROOT"
   

5. Clean up the temp file.

All sub-agent stamps land in <project>/reviews/INDEX.md with Trigger=pre-submission-report and roughly the same Last Run timestamp.

Why the orchestrator stamps (not the sub-agent)

Agents have inconsistent Bash tool grants at runtime (the 2026-05-19 harness investigation). The orchestrator always has Bash and always runs after the agents return. Moving stamping here decouples it from agent tool-surface uncertainty and fixes the burying problem (referee2-reviewer's stamping section used to live at line ~475 of a 518-line agent definition; agents reliably forgot to reach it).

Schema: ~/Task-Management/docs/reference/review-state-schema.md. Stamp directive format: ~/Task-Management/skills/_shared/stamp-directive-spec.md.