Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Tactic: Generate candidate changes, detect flip-points where conclusion reverses, measure fragility as distance to nearest flip.
.claude/skills/yogsoth-ai-minimal-change-search/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -3% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-19 | ✗→✓ | ▲ Improved | -14% | 0% |
Find the smallest perturbation that flips the conclusion — closer flip-points mean higher fragility.
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | causal-claim-extraction | Extract all causal claims (X causes Y, X leads to Y, X enables Y) from an artifact, producing a structured list of cause-effect pairs. | | counterfactual-scenario-construction | Construct precise, internally consistent counterfactual scenarios where specified factors are altered, then reason about the resulting conclusion. | | factor-enumeration | List all key factors, conditions, and assumptions that support or enable the artifact's conclusion. | | flip-point-detection | Find the minimal change magnitude along a dimension that causes the conclusion to flip from true to false. | | fragility-measurement | Compute a fragility index from flip-point distances and degradation scores, summarizing how robust the conclusion is. |
<!-- 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 | 43,229 | 37,628 | -13% | 1 | 1 | 0% | 5,707 | 6,847 | +20% | 0 | 0 | — |
case-02 | fail→fail | 21,263 | 23,360 | +10% | 1 | 1 | 0% | 2,822 | 5,268 | +87% | 0 | 0 | — |
case-03 | pass→fail | 25,813 | 32,084 | +24% | 1 | 1 | 0% | 3,826 | 6,870 | +80% | 0 | 0 | — |
case-04 | pass→fail | 35,722 | 40,584 | +14% | 1 | 1 | 0% | 7,303 | 8,388 | +15% | 0 | 0 | — |
case-05 | pass→fail | 26,870 | 42,833 | +59% | 1 | 1 | 0% | 3,935 | 7,620 | +94% | 0 | 0 | — |
case-06 | pass→pass | 20,681 | 5,562 | -73% | 1 | 1 | 0% | 2,457 | 1,518 | -38% | 0 | 0 | — |
case-07 | fail→pass | 14,763 | 15,390 | +4% | 1 | 1 | 0% | 2,528 | 2,460 | -3% | 0 | 0 | — |
case-08 | pass→pass | 18,852 | 12,702 | -33% | 1 | 1 | 0% | 1,923 | 1,750 | -9% | 0 | 0 | — |
case-09 | pass→pass | 23,532 | 12,802 | -46% | 1 | 1 | 0% | 2,721 | 1,789 | -34% | 0 | 0 | — |
case-10 | fail→pass | 21,623 | 7,291 | -66% | 1 | 1 | 0% | 1,686 | 862 | -49% | 0 | 0 | — |
case-11 | pass→pass | 13,790 | 8,350 | -39% | 1 | 1 | 0% | 1,398 | 1,152 | -18% | 0 | 0 | — |
case-12 | pass→pass | 13,403 | 7,744 | -42% | 1 | 1 | 0% | 1,274 | 930 | -27% | 0 | 0 | — |
case-13 | pass→pass | 31,843 | 7,756 | -76% | 1 | 1 | 0% | 1,859 | 943 | -49% | 0 | 0 | — |
case-14 | fail→fail | 49,346 | 17,073 | -65% | 1 | 1 | 0% | 1,115 | 1,870 | +68% | 0 | 0 | — |
case-15 | fail→pass | 20,202 | 18,479 | -9% | 1 | 1 | 0% | 2,652 | 2,757 | +4% | 0 | 0 | — |
case-16 | pass→pass | 16,465 | 8,812 | -46% | 1 | 1 | 0% | 2,847 | 2,137 | -25% | 0 | 0 | — |
case-17 | pass→pass | 18,511 | 13,927 | -25% | 1 | 1 | 0% | 2,210 | 1,860 | -16% | 0 | 0 | — |
case-18 | fail→pass | 17,696 | 14,908 | -16% | 1 | 1 | 0% | 2,040 | 2,130 | +4% | 0 | 0 | — |
case-19 | fail→pass | 20,478 | 19,160 | -6% | 1 | 1 | 0% | 2,379 | 2,049 | -14% | 0 | 0 | — |
case-20 | fail→pass | 13,835 | 6,998 | -49% | 1 | 1 | 0% | 2,020 | 1,592 | -21% | 0 | 0 | — |
case-21 | pass→pass | 15,797 | 4,166 | -74% | 1 | 1 | 0% | 1,647 | 1,296 | -21% | 0 | 0 | — |
case-22 | pass→pass | 4,229 | 8,651 | +105% | 1 | 1 | 0% | 769 | 1,155 | +50% | 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 +14 percentage points is the difference between those two pass rates over the 21 comparable cases. 3 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.