Verify that every quantitative claim in the paper is traceable to an analysis output file, and that no important output was omitted. Make sure to use this skill whenever the user wants to check that the paper and analysis are consistent before submission. Triggers include: "run the quality gate", "check the paper matches the analysis", "verify consistency", "does the paper match my results", "check my numbers", "are my tables right", "quality check before submission", "verify my claims", "make sure everything is consistent", "double-check the paper against my output files", or any pre-submission integrity check between paper text and computed results.
Cross-check every numerical claim in the paper against analysis output files. Reports only — never edits.
Input: $ARGUMENTS — path to the paper draft, or leave blank to auto-detect.
If $ARGUMENTS is provided, use that path. Otherwise glob for:
manuscripts/**/*.texmanuscripts/**/*.qmdmanuscripts/**/*.mdIf multiple drafts found, use AskUserQuestion to let the user pick:
Read the full manuscript and extract every quantitative claim:
Record location (section, paragraph, line number if available) for each claim.
Glob for all output files:
output/tables/**/*.tex — regression and summary tablesoutput/tables/**/*.html — HTML versionsoutput/figures/**/*.pdf, output/figures/**/*.png — figuresoutput/**/*.rds, output/**/*.pkl, output/**/*.parquet, output/**/*.csv — saved objectsBuild an inventory with file paths and sizes.
Dispatch the verifier agent via Task to perform the heavy verification work. Pass it:
Task prompt: "You are the verifier agent. Paper draft: [path].
Bibliography: [bib path].
CLAIMS TO VERIFY:
[paste the full claims list from Step 2]
OUTPUT FILE INVENTORY:
[paste the inventory from Step 3]
Verify each claim against the output files. Then do a reverse check —
find output files NOT referenced in the paper. Then check all citation
keys against the bibliography. Follow the verifier agent instructions
and return your full verification report."
After the verifier completes, collect its results:
Save to quality_reports/quality_gate_[YYYY-MM-DD]_[paper-name].md:
# Quality Gate Report: [Paper Name]
**Date:** [YYYY-MM-DD]
**Paper:** [file path]
## Verdict: PASS / CONDITIONAL PASS / FAIL
PASS = all claims matched, no missing citations, no unexplained unreferenced outputs
CONDITIONAL PASS = minor unverified claims or informational unreferenced outputs
FAIL = unverified critical claims or missing citations
---
## Claim Verification
| Claim | Location | Found in Output? | Source File | Status |
|-------|----------|-----------------|-------------|--------|
| β = 0.23 (SE = 0.04) | Section 4, para 2 | Yes | output/tables/main_regs.tex | MATCHED |
| N = 4,521 | Table 2 note | Yes | output/tables/main_regs.tex | MATCHED |
| 42% of firms | Intro, para 1 | No | — | UNVERIFIED |
---
## Unreferenced Outputs
Files in output/ not referenced in the paper:
| File | Size | Recommended Action |
|------|------|-------------------|
| output/tables/robustness_het.tex | 4.2 KB | Reference in Section 7 or explain exclusion |
---
## Missing Citations
| Key | Used At | Status |
|-----|---------|--------|
| SmithJones2021 | Section 3, para 1 | NOT IN BIBLIOGRAPHY — CRITICAL |
---
## Summary
- Claims verified: N / M total
- Claims unverified: K (see table above)
- Unreferenced outputs: J
- Missing citations: L
## Recommended Actions (Priority Order)
1. [BLOCKING] ...
2. [RECOMMENDED] ...