Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Strategy for large-N sparse pairwise comparison using TrueSkill, active learning, and rank centrality to rank 100+ candidates from limited comparisons.
.claude/skills/yogsoth-ai-efficient-exploration/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -10% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 19% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -46% | 0% |
Produce reliable rankings when the candidate set is too large for complete comparison. Uses information-theoretic pair selection and sparse-matrix rating algorithms to converge quickly with minimal comparisons.
| Resource | Allocation | |----------|-----------| | Comparisons | N×log(N) to 3N×log(N) | | Iterations | 5-20 rounds of adaptive selection | | Convergence target | Top-k stability ≥ 90% for 3 consecutive rounds |
yamlcandidates: [] # full candidate list comparison_history: [] # [{pair, winner, confidence, round}] ratings: {} # candidate → {mu, sigma} method: "" # trueskill | bt-incomplete | rank-centrality iteration: 0 budget_remaining: 0 convergence: {stable: false, score: 0.0, top_k_stable: false}
yamlranking: - {rank: 1, candidate: "...", mu: 38.2, sigma: 1.4, ci: [35.4, 41.0]} - {rank: 2, candidate: "...", mu: 36.8, sigma: 1.6, ci: [33.6, 40.0]} method: trueskill total_comparisons: 847 budget_utilization: 0.92 top_10_stability: 0.96 convergence_round: 14
<!-- BEGIN available-tables (generated) -->
Optional, no fixed order; the final leaf is always a sop.
| Tactic | When to use | | --- | --- | | adaptive-pair-selection | Iteratively select maximally informative pairs, execute comparisons, update ratings, and check convergence until ranking stabilizes. | | consistency-audit-loop | Detect preference cycles, localize inconsistent judgments, request corrections, and recompute ratings until consistency threshold is met. |
Optional, no fixed order; the final leaf is always a sop.
| SOP | When to use | | --- | --- | | ranking-synthesis | Produce the final ranking artifact from converged ratings and consistency report. |
<!-- 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 | 22,209 | 11,904 | -46% | 1 | 1 | 0% | 4,586 | 3,116 | -32% | 0 | 0 | — |
case-02 | fail→pass | 18,026 | 10,704 | -41% | 1 | 1 | 0% | 3,256 | 2,934 | -10% | 0 | 0 | — |
case-03 | fail→pass | 22,240 | 14,456 | -35% | 1 | 1 | 0% | 4,527 | 3,610 | -20% | 0 | 0 | — |
case-13 | fail→pass | 19,469 | 10,435 | -46% | 1 | 1 | 0% | 2,995 | 2,399 | -20% | 0 | 0 | — |
case-04 | pass→pass | 17,604 | 14,310 | -19% | 1 | 1 | 0% | 2,750 | 3,286 | +19% | 0 | 0 | — |
case-05 | pass→pass | 12,548 | 14,079 | +12% | 1 | 1 | 0% | 2,171 | 3,023 | +39% | 0 | 0 | — |
case-06 | pass→fail | 15,877 | 18,354 | +16% | 1 | 1 | 0% | 2,763 | 3,563 | +29% | 0 | 0 | — |
case-07 | pass→pass | 13,616 | 4,677 | -66% | 1 | 1 | 0% | 2,412 | 1,559 | -35% | 0 | 0 | — |
case-08 | fail→pass | 24,581 | 5,299 | -78% | 1 | 1 | 0% | 1,417 | 1,691 | +19% | 0 | 0 | — |
case-09 | fail→pass | 13,277 | 2,366 | -82% | 1 | 1 | 0% | 2,124 | 1,155 | -46% | 0 | 0 | — |
case-10 | fail→pass | 24,757 | 2,139 | -91% | 1 | 1 | 0% | 1,152 | 1,068 | -7% | 0 | 0 | — |
case-11 | fail→pass | 12,974 | 2,857 | -78% | 1 | 1 | 0% | 2,325 | 1,266 | -46% | 0 | 0 | — |
case-12 | pass→pass | 17,232 | 3,959 | -77% | 1 | 1 | 0% | 2,606 | 1,334 | -49% | 0 | 0 | — |
case-14 | fail→pass | 30,142 | 6,992 | -77% | 1 | 1 | 0% | 1,173 | 1,824 | +55% | 0 | 0 | — |
case-15 | fail→pass | 9,256 | 3,940 | -57% | 1 | 1 | 0% | 1,692 | 1,497 | -12% | 0 | 0 | — |
case-16 | fail→pass | 4,816 | 2,836 | -41% | 1 | 1 | 0% | 669 | 1,287 | +92% | 0 | 0 | — |
case-17 | fail→pass | 26,076 | 1,932 | -93% | 1 | 1 | 0% | 1,138 | 1,057 | -7% | 0 | 0 | — |
case-18 | pass→pass | 18,203 | 1,799 | -90% | 1 | 1 | 0% | 1,406 | 1,047 | -26% | 0 | 0 | — |
case-19 | fail→pass | 13,692 | 2,885 | -79% | 1 | 1 | 0% | 2,119 | 1,217 | -43% | 0 | 0 | — |
case-20 | fail→pass | 10,799 | 2,470 | -77% | 1 | 1 | 0% | 1,590 | 1,068 | -33% | 0 | 0 | — |
case-21 | fail→pass | 11,927 | 1,768 | -85% | 1 | 1 | 0% | 1,892 | 1,085 | -43% | 0 | 0 | — |
case-22 | pass→pass | 14,861 | 3,850 | -74% | 1 | 1 | 0% | 2,503 | 1,379 | -45% | 0 | 0 | — |
case-23 | fail→pass | 9,337 | 2,363 | -75% | 1 | 1 | 0% | 1,606 | 1,162 | -28% | 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. 23 cases were attempted, and 19 counted toward the lift figure. The other 4 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 +61 percentage points is the difference between those two pass rates over the 19 comparable cases. 1 case got worse with the skill loaded, and it is 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.