Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →One-at-a-time assumption perturbation — extract assumptions, define negations, re-derive conclusions under each negation, measure sensitivity. Identifies which assumptions are load-bearing.
.claude/skills/yogsoth-ai-deep-insight-assumption-perturbation/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 97% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 110% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-09 | ✗→✓ | ▲ Improved | 90% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 84% | 0% |
Systematically perturb assumptions to identify which are load-bearing.
assumption-extraction → negation-definition → re-derivation → conclusion-sensitivity-measurement
Subagent: assumption-extraction, negation-definition, re-derivation, conclusion-sensitivity-measurement Shared: assumption-surfacing Import: paper-research
Extract all assumptions (use shared SOP for initial surfacing), define weakest plausible alternative for each, re-derive the conclusion under each alternative, measure change magnitude and direction. Rank by sensitivity.
Key principle: negation is not logical NOT — it is the strongest plausible alternative. "Data is normally distributed" negates to "data follows a heavy-tailed distribution" not "data is not normally distributed."
<HARD-GATE>
- Assumptions extracted: >= 5
- Negations defined: >= 5
- Re-derivations completed: >= 4
- Sensitivity rankings produced: >= 1 complete ranking
</HARD-GATE><!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | conclusion-sensitivity-measurement | Quantify how much conclusions change across all assumption negations and produce a sensitivity ranking. | | deep-insight-assumption-extraction | Systematically extract all assumptions (stated, implicit, boundary, mathematical, practical) from a method or model. | | deep-insight-assumption-surfacing | Systematically extract implicit assumptions from methods, frameworks, or arguments. Identifies what is taken for granted without explicit justification. | | deep-insight-paper-research | Full-text paper reading via three-pass Keshav method. Import of literature-engine/literature-research skill. Authoritative source for claims about paper content. | | negation-definition | Define strongest plausible alternatives (negations) for each assumption to enable perturbation analysis. | | re-derivation | Re-derive conclusions under a negated assumption, tracking where the derivation diverges from the original. |
<!-- 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-02 | pass→pass | 37,046 | 31,015 | -16% | 1 | 1 | 0% | 4,070 | 6,666 | +64% | 0 | 0 | — |
case-01 | fail→fail | 21,598 | 32,295 | +50% | 1 | 1 | 0% | 3,589 | 6,672 | +86% | 0 | 0 | — |
case-03 | fail→pass | 17,686 | 48,720 | +175% | 1 | 1 | 0% | 2,744 | 5,417 | +97% | 0 | 0 | — |
case-04 | pass→pass | 20,981 | 96,982 | +362% | 1 | 1 | 0% | 3,459 | 6,648 | +92% | 0 | 0 | — |
case-05 | fail→pass | 20,439 | 34,481 | +69% | 1 | 1 | 0% | 3,154 | 6,638 | +110% | 0 | 0 | — |
case-06 | pass→pass | 18,260 | 34,741 | +90% | 1 | 1 | 0% | 2,706 | 6,646 | +146% | 0 | 0 | — |
case-07 | fail→pass | 26,223 | 35,385 | +35% | 1 | 1 | 0% | 4,756 | 6,512 | +37% | 0 | 0 | — |
case-08 | pass→pass | 25,648 | 31,851 | +24% | 1 | 1 | 0% | 4,902 | 6,652 | +36% | 0 | 0 | — |
case-09 | fail→pass | 32,221 | 10,054 | -69% | 1 | 1 | 0% | 1,128 | 2,138 | +90% | 0 | 0 | — |
case-10 | pass→pass | 26,349 | 25,564 | -3% | 1 | 1 | 0% | 3,001 | 4,969 | +66% | 0 | 0 | — |
case-11 | pass→pass | 20,448 | 37,157 | +82% | 1 | 1 | 0% | 3,305 | 6,640 | +101% | 0 | 0 | — |
case-12 | pass→pass | 20,127 | 38,293 | +90% | 1 | 1 | 0% | 3,109 | 6,633 | +113% | 0 | 0 | — |
case-13 | fail→fail | 17,245 | 35,150 | +104% | 1 | 1 | 0% | 2,780 | 6,635 | +139% | 0 | 0 | — |
case-14 | fail→pass | 23,089 | 36,839 | +60% | 1 | 1 | 0% | 3,599 | 6,633 | +84% | 0 | 0 | — |
case-15 | pass→pass | 20,859 | 33,136 | +59% | 1 | 1 | 0% | 3,413 | 6,636 | +94% | 0 | 0 | — |
case-16 | fail→pass | 15,273 | 10,338 | -32% | 1 | 1 | 0% | 2,308 | 2,070 | -10% | 0 | 0 | — |
case-17 | fail→pass | 20,117 | 33,874 | +68% | 1 | 1 | 0% | 3,254 | 6,633 | +104% | 0 | 0 | — |
case-18 | fail→pass | 20,725 | 36,189 | +75% | 1 | 1 | 0% | 3,173 | 6,250 | +97% | 0 | 0 | — |
case-19 | fail→fail | 18,332 | 35,435 | +93% | 1 | 1 | 0% | 3,092 | 6,630 | +114% | 0 | 0 | — |
case-20 | pass→pass | 23,072 | 32,253 | +40% | 1 | 1 | 0% | 4,271 | 6,600 | +55% | 0 | 0 | — |
case-21 | pass→pass | 22,901 | 29,481 | +29% | 1 | 1 | 0% | 4,367 | 6,677 | +53% | 0 | 0 | — |
case-22 | fail→fail | 2,674 | 44,448 | +1562% | 1 | 1 | 0% | 336 | 6,206 | +1747% | 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 +36 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.