Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when the user wants to verify paper claims against code or data, audit numerical accuracy, check formula-code alignment, or validate citation accuracy. Triggers on phrases like "verify claims", "check numbers", "do the numbers match", "formula vs code", "audit the paper", or "cross-check results".
.claude/skills/fcakyon-paper-verification/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 19% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 20% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 52% | 0% |
| case-16 | ✗→✓ | ▲ Improved | 21% | 0% |
You are helping a researcher verify that their paper accurately reflects their code and experimental results. This is the most critical quality control step in academic writing.
For every number in the paper (dataset sizes, metric values, percentages, counts):
Template:
| Paper claim | Location (.tex) | Source file/code | Source value | Match? |
|-------------|-----------------|-----------------|-------------|--------|
| "13,999 frames" | abstract L3 | len(glob(labels/*.json)) | ? | ? |
| "4.2% improvement" | Table 2 | eval_results.json | ? | ? |Common numerical errors:
For each method described in the paper:
Common mismatches:
For each equation in the paper:
For each citation in the paper:
Step 1: Extract the claim and the cited paper Step 2: Verify BibTeX metadata against DBLP:
Step 3: For cited claims with specific numbers:
Step 4: Check for common citation errors:
Produce a structured verification report:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-08 | pass→pass | 8,074 | 4,913 | -39% | 1 | 1 | 0% | 1,262 | 1,807 | +43% | 0 | 0 | — |
case-01 | fail→pass | 85,194 | 8,864 | -90% | 1 | 1 | 0% | 5,083 | 2,592 | -49% | 0 | 0 | — |
case-02 | fail→fail | 9,117 | 12,173 | +34% | 1 | 1 | 0% | 1,512 | 3,252 | +115% | 0 | 0 | — |
case-03 | fail→fail | 7,965 | 8,746 | +10% | 1 | 1 | 0% | 1,417 | 2,572 | +82% | 0 | 0 | — |
case-04 | pass→pass | 9,475 | 5,819 | -39% | 1 | 1 | 0% | 1,467 | 1,936 | +32% | 0 | 0 | — |
case-05 | fail→pass | 8,924 | 4,493 | -50% | 1 | 1 | 0% | 1,439 | 1,711 | +19% | 0 | 0 | — |
case-06 | pass→pass | 8,376 | 4,196 | -50% | 1 | 1 | 0% | 1,152 | 1,656 | +44% | 0 | 0 | — |
case-07 | pass→pass | 15,231 | 9,780 | -36% | 1 | 1 | 0% | 2,399 | 2,569 | +7% | 0 | 0 | — |
case-09 | fail→pass | 12,540 | 7,587 | -39% | 1 | 1 | 0% | 1,908 | 2,283 | +20% | 0 | 0 | — |
case-10 | pass→pass | 6,989 | 4,389 | -37% | 1 | 1 | 0% | 1,146 | 1,701 | +48% | 0 | 0 | — |
case-11 | pass→pass | 9,675 | 5,275 | -45% | 1 | 1 | 0% | 1,497 | 1,856 | +24% | 0 | 0 | — |
case-12 | pass→pass | 8,627 | 5,220 | -39% | 1 | 1 | 0% | 1,441 | 1,850 | +28% | 0 | 0 | — |
case-21 | pass→pass | 11,748 | 7,949 | -32% | 1 | 1 | 0% | 1,998 | 2,357 | +18% | 0 | 0 | — |
case-13 | pass→pass | 9,486 | 6,259 | -34% | 1 | 1 | 0% | 1,384 | 1,955 | +41% | 0 | 0 | — |
case-14 | pass→pass | 9,060 | 5,649 | -38% | 1 | 1 | 0% | 1,405 | 1,928 | +37% | 0 | 0 | — |
case-15 | fail→pass | 7,722 | 4,934 | -36% | 1 | 1 | 0% | 1,206 | 1,837 | +52% | 0 | 0 | — |
case-16 | fail→pass | 15,067 | 10,151 | -33% | 1 | 1 | 0% | 2,282 | 2,756 | +21% | 0 | 0 | — |
case-17 | pass→pass | 7,189 | 4,207 | -41% | 1 | 1 | 0% | 1,031 | 1,604 | +56% | 0 | 0 | — |
case-18 | pass→pass | 4,705 | 7,275 | +55% | 1 | 1 | 0% | 878 | 2,166 | +147% | 0 | 0 | — |
case-19 | fail→pass | 12,074 | 2,834 | -77% | 1 | 1 | 0% | 1,953 | 1,431 | -27% | 0 | 0 | — |
case-20 | pass→pass | 10,146 | 9,569 | -6% | 1 | 1 | 0% | 1,496 | 2,368 | +58% | 0 | 0 | — |
case-22 | pass→pass | 15,886 | 10,143 | -36% | 1 | 1 | 0% | 3,125 | 3,078 | -2% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +27 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.