Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →(Proposal, unverified) Attempt to verify a paper's reported results by actually executing its released code/scripts against its own reported configuration — the only SOP in this package whose action type is code execution rather than text reading/judgment. Use this after unit-classification has extracted the paper's reported configuration/hyperparameters as classified units; "not_attempted" is a correct, common output when the paper's own reporting is too incomplete to run, not a failure of this
.claude/skills/yogsoth-ai-reproducibility-third-party-verification/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 83% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -5% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 34% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -6% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -52% | 0% |
Actually runs code to check reported results against the paper's own extracted configuration — unique action type (execution) in this package. Fills the evidence-verification × engineering-metadata gap in the evaluative-stance × content-layer matrix.
Subagent — spawned via spawn-agent skill.
This SOP needs the paper's reported configuration already pulled out in structured form before attempting to verify it — hence its input is classified_units, not full_text directly. An earlier graph draft had this SOP depending on nothing upstream, which meant it had no defined way to get the structured claims it needs to check.
No primary-source precedent, no inter-rater-reliability baseline. Keep "(Proposal, unverified)" in the description until real usage validates the method. Given the code-execution action type, treat any scope expansion here with more caution than the other 3 proposal SOPs.
<!-- BEGIN available-tables (generated) -->
| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 9,642 | 11,077 | +15% | 1 | 1 | 0% | 633 | 901 | +42% | 0 | 0 | — |
case-02 | fail→fail | 9,774 | 7,984 | -18% | 1 | 1 | 0% | 742 | 712 | -4% | 0 | 0 | — |
case-03 | fail→fail | 14,563 | 22,848 | +57% | 1 | 1 | 0% | 1,469 | 3,281 | +123% | 0 | 0 | — |
case-04 | fail→pass | 10,504 | 12,803 | +22% | 1 | 1 | 0% | 856 | 1,568 | +83% | 0 | 0 | — |
case-05 | fail→fail | 15,321 | 11,705 | -24% | 1 | 1 | 0% | 1,619 | 1,041 | -36% | 0 | 0 | — |
case-06 | fail→pass | 13,799 | 12,021 | -13% | 1 | 1 | 0% | 1,562 | 1,482 | -5% | 0 | 0 | — |
case-07 | fail→pass | 40,638 | 12,802 | -68% | 1 | 1 | 0% | 1,201 | 1,607 | +34% | 0 | 0 | — |
case-08 | fail→pass | 18,128 | 17,405 | -4% | 1 | 1 | 0% | 2,511 | 2,361 | -6% | 0 | 0 | — |
case-09 | pass→pass | 13,018 | 8,948 | -31% | 1 | 1 | 0% | 1,304 | 949 | -27% | 0 | 0 | — |
case-10 | fail→fail | 18,755 | 11,039 | -41% | 1 | 1 | 0% | 2,282 | 1,260 | -45% | 0 | 0 | — |
case-11 | pass→pass | 17,363 | 13,679 | -21% | 1 | 1 | 0% | 1,839 | 1,719 | -7% | 0 | 0 | — |
case-12 | pass→pass | 20,086 | 9,879 | -51% | 1 | 1 | 0% | 2,125 | 1,053 | -50% | 0 | 0 | — |
case-13 | fail→pass | 13,257 | 7,203 | -46% | 1 | 1 | 0% | 1,238 | 598 | -52% | 0 | 0 | — |
case-14 | fail→pass | 19,020 | 12,116 | -36% | 1 | 1 | 0% | 1,869 | 1,531 | -18% | 0 | 0 | — |
case-15 | pass→pass | 20,792 | 16,676 | -20% | 1 | 1 | 0% | 2,439 | 2,008 | -18% | 0 | 0 | — |
case-16 | fail→pass | 23,644 | 15,525 | -34% | 1 | 1 | 0% | 2,875 | 1,886 | -34% | 0 | 0 | — |
case-17 | fail→pass | 17,887 | 7,946 | -56% | 1 | 1 | 0% | 1,788 | 756 | -58% | 0 | 0 | — |
case-18 | pass→pass | 20,399 | 19,740 | -3% | 1 | 1 | 0% | 2,320 | 2,550 | +10% | 0 | 0 | — |
case-19 | fail→pass | 19,062 | 7,611 | -60% | 1 | 1 | 0% | 2,060 | 654 | -68% | 0 | 0 | — |
case-20 | fail→fail | 12,313 | 23,576 | +91% | 1 | 1 | 0% | 1,410 | 3,823 | +171% | 0 | 0 | — |
case-21 | pass→pass | 28,263 | 30,218 | +7% | 1 | 1 | 0% | 3,796 | 2,720 | -28% | 0 | 0 | — |
case-22 | pass→pass | 14,323 | 12,653 | -12% | 1 | 1 | 0% | 1,817 | 1,616 | -11% | 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, and 21 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +41 percentage points is the difference between those two pass rates over the 21 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.