Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Test conclusion robustness via multi-model convergence — enumerate assumptions, generate alternatives, compare results, flag fragile conclusions.
.claude/skills/yogsoth-ai-robustness-testing/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -69% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 14% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 31% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 7% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 41% | 0% |
Test whether conclusions survive across different modeling choices.
| Base SOP | Target | ±10% Range | |----------|--------|------------| | web-search | 20 | 18–22 | | web-research | 10 | 9–11 | | paper-overview | 30 | 27–33 | | paper-search | 25 | 22–28 | | paper-research | 15 | 13–17 |
<HARD-GATE>
| SOP | Done | Target | % |
|-----|------|--------|---|
| web-search | ? | 20 | ? |
| web-research | ? | 10 | ? |
| paper-overview | ? | 30 | ? |
| paper-search | ? | 25 | ? |
| paper-research | ? | 15 | ? |
Budget Gate: OPEN/CLOSED (>=80% required to exit)
</HARD-GATE>Import: web-search, web-research, paper-overview, paper-search, paper-research Subagent: assumption-enumeration, alternative-model-generation, convergence-assessment, fragility-flagging
Enumerate modeling assumptions, generate alternative models by relaxing each, compare results across alternatives, flag results that depend on specific assumptions (fragile).
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | multi-model-convergence | Wimsatt-style multi-method cross-validation — enumerate assumptions, generate alternative models, compare results, flag divergences. |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | alternative-model-generation | Generate alternative model formulations by relaxing, replacing, or generalizing specific assumptions. | | convergence-assessment | Compare results across multiple model variants — quantitative agreement metrics and qualitative conclusion stability. | | deep-insight-assumption-enumeration | Systematically identify all assumptions in a method/model — structural, parametric, distributional, and scope assumptions. | | fragility-flagging | Identify which specific assumption changes cause conclusion divergence. Rates fragility severity and plausibility of alternatives. |
<!-- 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→pass | 47,203 | 15,783 | -67% | 1 | 1 | 0% | 7,027 | 2,179 | -69% | 0 | 0 | — |
case-02 | fail→pass | 49,065 | 48,426 | -1% | 1 | 1 | 0% | 7,756 | 8,832 | +14% | 0 | 0 | — |
case-03 | fail→pass | 43,621 | 51,875 | +19% | 1 | 1 | 0% | 6,763 | 8,828 | +31% | 0 | 0 | — |
case-04 | pass→fail | 18,223 | 22,954 | +26% | 1 | 1 | 0% | 2,197 | 1,772 | -19% | 0 | 0 | — |
case-05 | pass→fail | 15,200 | 47,267 | +211% | 1 | 1 | 0% | 2,087 | 8,823 | +323% | 0 | 0 | — |
case-06 | pass→pass | 29,448 | 51,292 | +74% | 1 | 1 | 0% | 4,282 | 8,795 | +105% | 0 | 0 | — |
case-07 | fail→pass | 49,087 | 51,896 | +6% | 1 | 1 | 0% | 8,253 | 8,821 | +7% | 0 | 0 | — |
case-08 | fail→fail | 49,137 | 46,574 | -5% | 1 | 1 | 0% | 7,653 | 8,813 | +15% | 0 | 0 | — |
case-09 | fail→fail | 47,043 | 49,208 | +5% | 1 | 1 | 0% | 8,238 | 8,806 | +7% | 0 | 0 | — |
case-10 | fail→fail | 17,825 | 13,066 | -27% | 1 | 1 | 0% | 2,035 | 2,123 | +4% | 0 | 0 | — |
case-11 | fail→pass | 27,874 | 43,180 | +55% | 1 | 1 | 0% | 4,989 | 7,021 | +41% | 0 | 0 | — |
case-12 | fail→pass | 46,032 | 50,588 | +10% | 1 | 1 | 0% | 7,615 | 8,795 | +15% | 0 | 0 | — |
case-13 | fail→fail | 40,684 | 53,444 | +31% | 1 | 1 | 0% | 6,488 | 8,794 | +36% | 0 | 0 | — |
case-14 | fail→fail | 80,817 | 51,281 | -37% | 1 | 1 | 0% | 7,688 | 8,799 | +14% | 0 | 0 | — |
case-15 | fail→pass | 19,316 | 10,336 | -46% | 1 | 1 | 0% | 2,497 | 1,794 | -28% | 0 | 0 | — |
case-16 | fail→pass | 24,296 | 49,997 | +106% | 1 | 1 | 0% | 3,292 | 8,809 | +168% | 0 | 0 | — |
case-17 | fail→pass | 39,310 | 48,764 | +24% | 1 | 1 | 0% | 7,039 | 8,789 | +25% | 0 | 0 | — |
case-18 | fail→pass | 45,946 | 51,503 | +12% | 1 | 1 | 0% | 8,221 | 8,789 | +7% | 0 | 0 | — |
case-19 | fail→fail | 41,708 | 66,854 | +60% | 1 | 1 | 0% | 6,508 | 8,795 | +35% | 0 | 0 | — |
case-20 | fail→pass | 31,128 | 43,129 | +39% | 1 | 1 | 0% | 6,516 | 8,794 | +35% | 0 | 0 | — |
case-21 | fail→pass | 34,643 | 53,677 | +55% | 1 | 1 | 0% | 5,226 | 8,796 | +68% | 0 | 0 | — |
case-22 | fail→fail | 45,844 | 51,704 | +13% | 1 | 1 | 0% | 8,232 | 8,800 | +7% | 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 +45 percentage points is the difference between those two pass rates over the 21 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
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.