Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Identify which specific comparison pairs are most responsible for preference cycles and inconsistencies.
.claude/skills/yogsoth-ai-inconsistency-localization/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -75% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -78% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -48% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -77% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -48% | 0% |
Given detected cycles in a preference graph, identifies which specific pairwise judgments are most likely erroneous. Ranks problematic pairs by their participation in cycles and weakness of evidence.
Runs as a subagent. Receives the comparison matrix and detected cycles, returns prioritized list of pairs to re-evaluate.
Localization requires cross-referencing cycle membership with edge confidence scores and computing centrality metrics on the inconsistency subgraph. This focused analysis benefits from isolation.
Output MUST contain at least one problematic pair if cycles input is non-empty. Each pair MUST appear in at least one of the input cycles. Pairs MUST be ordered by priority (most problematic first).
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. Used by SOPs that declare execution: subagent. |
<!-- 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 | pass→pass | 21,408 | 55,250 | +158% | 1 | 1 | 0% | 1,900 | 1,001 | -47% | 0 | 0 | — |
case-02 | pass→pass | 48,091 | 43,561 | -9% | 1 | 1 | 0% | 1,308 | 1,181 | -10% | 0 | 0 | — |
case-03 | pass→pass | 34,685 | 70,465 | +103% | 1 | 1 | 0% | 3,278 | 2,516 | -23% | 0 | 0 | — |
case-04 | fail→pass | 79,546 | 40,929 | -49% | 1 | 1 | 0% | 3,674 | 930 | -75% | 0 | 0 | — |
case-05 | pass→pass | 11,079 | 29,220 | +164% | 1 | 1 | 0% | 1,131 | 1,387 | +23% | 0 | 0 | — |
case-06 | pass→pass | 24,489 | 13,822 | -44% | 1 | 1 | 0% | 2,592 | 1,674 | -35% | 0 | 0 | — |
case-07 | pass→pass | 13,365 | 7,232 | -46% | 1 | 1 | 0% | 1,355 | 1,127 | -17% | 0 | 0 | — |
case-08 | fail→pass | 16,201 | 8,966 | -45% | 1 | 1 | 0% | 2,293 | 511 | -78% | 0 | 0 | — |
case-09 | pass→pass | 14,054 | 10,516 | -25% | 1 | 1 | 0% | 1,925 | 1,431 | -26% | 0 | 0 | — |
case-10 | pass→pass | 19,186 | 9,462 | -51% | 1 | 1 | 0% | 1,176 | 1,090 | -7% | 0 | 0 | — |
case-11 | pass→pass | 9,271 | 9,693 | +5% | 1 | 1 | 0% | 1,583 | 908 | -43% | 0 | 0 | — |
case-12 | fail→pass | 22,016 | 9,431 | -57% | 1 | 1 | 0% | 1,867 | 967 | -48% | 0 | 0 | — |
case-13 | fail→pass | 28,665 | 7,807 | -73% | 1 | 1 | 0% | 2,218 | 514 | -77% | 0 | 0 | — |
case-14 | fail→pass | 17,979 | 20,624 | +15% | 1 | 1 | 0% | 2,992 | 1,569 | -48% | 0 | 0 | — |
case-15 | fail→pass | 14,631 | 7,717 | -47% | 1 | 1 | 0% | 2,502 | 1,259 | -50% | 0 | 0 | — |
case-16 | pass→pass | 17,521 | 7,183 | -59% | 1 | 1 | 0% | 1,935 | 753 | -61% | 0 | 0 | — |
case-17 | fail→fail | 11,492 | 3,398 | -70% | 1 | 1 | 0% | 2,019 | 768 | -62% | 0 | 0 | — |
case-18 | pass→pass | 45,886 | 18,636 | -59% | 1 | 1 | 0% | 2,731 | 2,477 | -9% | 0 | 0 | — |
case-19 | pass→pass | 15,246 | 10,844 | -29% | 1 | 1 | 0% | 1,747 | 1,504 | -14% | 0 | 0 | — |
case-20 | pass→pass | 16,219 | 30,125 | +86% | 1 | 1 | 0% | 3,308 | 4,446 | +34% | 0 | 0 | — |
case-21 | pass→pass | 15,404 | 19,094 | +24% | 1 | 1 | 0% | 2,684 | 3,456 | +29% | 0 | 0 | — |
case-22 | fail→pass | 21,868 | 12,015 | -45% | 1 | 1 | 0% | 2,235 | 2,194 | -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 +32 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.