Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Detect preference cycles, localize inconsistent judgments, request corrections, and recompute ratings until consistency threshold is met.
.claude/skills/yogsoth-ai-consistency-audit-loop/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-06 | ✗→✓ | ▲ Improved | 5% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -62% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -53% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -62% | 0% |
Audit a comparison matrix for transitivity violations, pinpoint the most problematic judgments, request re-evaluation of those pairs, and recompute until the consistency threshold is satisfied.
Loop stages 1-3 until consistency threshold met or repair budget exhausted.
| Stage | SOP | Input | Output | |-------|-----|-------|--------| | Detect | cycle-detection | comparison_matrix | cycles], transitivity_score | | Localize | inconsistency-localization | comparison_matrix, cycles] | problematic_pairs] | | Repair | comparison-executor | pair, context | judgment |
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | comparison-executor | Execute a pairwise comparison between two candidates, producing a judgment with winner, confidence, and reasoning. | | cycle-detection | Scan a pairwise comparison matrix for preference cycles and compute transitivity metrics. | | inconsistency-localization | Identify which specific comparison pairs are most responsible for preference cycles and inconsistencies. |
<!-- 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 | 31,259 | 31,423 | +1% | 1 | 1 | 0% | 6,244 | 6,748 | +8% | 0 | 0 | — |
case-02 | fail→fail | 26,553 | 30,559 | +15% | 1 | 1 | 0% | 6,236 | 6,740 | +8% | 0 | 0 | — |
case-03 | pass→pass | 7,654 | 3,893 | -49% | 1 | 1 | 0% | 1,343 | 1,088 | -19% | 0 | 0 | — |
case-04 | pass→pass | 8,939 | 2,687 | -70% | 1 | 1 | 0% | 1,316 | 903 | -31% | 0 | 0 | — |
case-05 | pass→pass | 15,098 | 5,723 | -62% | 1 | 1 | 0% | 2,144 | 1,413 | -34% | 0 | 0 | — |
case-06 | fail→pass | 5,085 | 2,703 | -47% | 1 | 1 | 0% | 908 | 952 | +5% | 0 | 0 | — |
case-07 | fail→pass | 11,966 | 2,079 | -83% | 1 | 1 | 0% | 2,123 | 814 | -62% | 0 | 0 | — |
case-08 | pass→pass | 14,088 | 5,839 | -59% | 1 | 1 | 0% | 2,161 | 1,475 | -32% | 0 | 0 | — |
case-09 | fail→pass | 11,792 | 2,701 | -77% | 1 | 1 | 0% | 1,960 | 930 | -53% | 0 | 0 | — |
case-10 | fail→pass | 14,700 | 2,335 | -84% | 1 | 1 | 0% | 2,281 | 867 | -62% | 0 | 0 | — |
case-11 | fail→pass | 13,019 | 2,647 | -80% | 1 | 1 | 0% | 2,008 | 944 | -53% | 0 | 0 | — |
case-12 | pass→pass | 13,282 | 4,373 | -67% | 1 | 1 | 0% | 2,234 | 1,187 | -47% | 0 | 0 | — |
case-13 | fail→fail | 16,585 | 5,523 | -67% | 1 | 1 | 0% | 2,454 | 1,358 | -45% | 0 | 0 | — |
case-14 | fail→pass | 11,635 | 2,114 | -82% | 1 | 1 | 0% | 1,897 | 812 | -57% | 0 | 0 | — |
case-15 | fail→pass | 11,625 | 3,381 | -71% | 1 | 1 | 0% | 1,676 | 1,019 | -39% | 0 | 0 | — |
case-16 | fail→pass | 25,005 | 3,187 | -87% | 1 | 1 | 0% | 1,015 | 982 | -3% | 0 | 0 | — |
case-17 | pass→pass | 12,492 | 3,553 | -72% | 1 | 1 | 0% | 1,892 | 1,108 | -41% | 0 | 0 | — |
case-18 | fail→pass | 11,366 | 2,385 | -79% | 1 | 1 | 0% | 1,801 | 855 | -53% | 0 | 0 | — |
case-19 | fail→pass | 14,210 | 2,731 | -81% | 1 | 1 | 0% | 1,937 | 871 | -55% | 0 | 0 | — |
case-20 | pass→pass | 20,495 | 29,154 | +42% | 1 | 1 | 0% | 3,993 | 6,681 | +67% | 0 | 0 | — |
case-21 | pass→pass | 18,106 | 31,920 | +76% | 1 | 1 | 0% | 3,471 | 6,675 | +92% | 0 | 0 | — |
case-22 | fail→fail | 6,141 | 10,322 | +68% | 1 | 1 | 0% | 1,230 | 2,413 | +96% | 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 +50 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.