Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →SOP for attaching evidence to a claim — create typed edge (supported_by, contradicts, qualifies) with evidence quality metadata.
.claude/skills/yogsoth-ai-evidence-attachment/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -42% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -67% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -56% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -30% | 0% |
Attach a piece of evidence to a claim with a typed relationship edge.
vault_add_edge + vault_search
supported_by — evidence directly supports the claimcontradicts — evidence directly contradicts the claimqualifies — evidence limits the scope or adds conditions (use related_to edge)edge_type: supported_by | contradicts | related_toevidence_quality: strong | moderate | weakdirectness: direct | indirect | analogical[[dir/slug]] pointing to the claim (dir/slug = target path minus .md). Place inline at semantically relevant location. Skip if already present.<HARD-GATE> Must specify evidence quality and directness. "Supported_by" without quality assessment is not allowed. </HARD-GATE>
Returns: { claim: string, evidence: string, edge_type: string, quality: string }
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 10,058 | 5,351 | -47% | 1 | 1 | 0% | 1,599 | 704 | -56% | 0 | 0 | — |
case-02 | fail→fail | 3,271 | 4,751 | +45% | 1 | 1 | 0% | 212 | 559 | +164% | 0 | 0 | — |
case-03 | fail→fail | 12,916 | 4,449 | -66% | 1 | 1 | 0% | 2,092 | 641 | -69% | 0 | 0 | — |
case-08 | pass→pass | 6,985 | 2,601 | -63% | 1 | 1 | 0% | 1,114 | 717 | -36% | 0 | 0 | — |
case-04 | fail→pass | 10,869 | 4,224 | -61% | 1 | 1 | 0% | 1,752 | 1,021 | -42% | 0 | 0 | — |
case-05 | fail→fail | 12,698 | 4,299 | -66% | 1 | 1 | 0% | 2,013 | 1,100 | -45% | 0 | 0 | — |
case-06 | pass→pass | 8,696 | 2,623 | -70% | 1 | 1 | 0% | 1,489 | 783 | -47% | 0 | 0 | — |
case-07 | pass→pass | 14,007 | 2,183 | -84% | 1 | 1 | 0% | 1,770 | 709 | -60% | 0 | 0 | — |
case-09 | fail→pass | 9,969 | 1,809 | -82% | 1 | 1 | 0% | 1,796 | 595 | -67% | 0 | 0 | — |
case-10 | fail→pass | 5,748 | 6,400 | +11% | 1 | 1 | 0% | 937 | 863 | -8% | 0 | 0 | — |
case-11 | fail→pass | 10,353 | 2,716 | -74% | 1 | 1 | 0% | 1,854 | 818 | -56% | 0 | 0 | — |
case-12 | fail→fail | 8,046 | 1,745 | -78% | 1 | 1 | 0% | 1,207 | 552 | -54% | 0 | 0 | — |
case-13 | fail→pass | 9,127 | 5,141 | -44% | 1 | 1 | 0% | 1,636 | 1,140 | -30% | 0 | 0 | — |
case-14 | pass→pass | 8,878 | 3,580 | -60% | 1 | 1 | 0% | 1,341 | 792 | -41% | 0 | 0 | — |
case-15 | pass→pass | 10,769 | 1,847 | -83% | 1 | 1 | 0% | 1,787 | 623 | -65% | 0 | 0 | — |
case-16 | pass→pass | 6,564 | 2,183 | -67% | 1 | 1 | 0% | 1,100 | 674 | -39% | 0 | 0 | — |
case-17 | fail→pass | 13,719 | 2,139 | -84% | 1 | 1 | 0% | 2,145 | 625 | -71% | 0 | 0 | — |
case-18 | fail→pass | 10,012 | 5,400 | -46% | 1 | 1 | 0% | 1,547 | 1,246 | -19% | 0 | 0 | — |
case-19 | fail→pass | 14,682 | 2,185 | -85% | 1 | 1 | 0% | 2,326 | 677 | -71% | 0 | 0 | — |
case-20 | pass→fail | 10,824 | 6,194 | -43% | 1 | 1 | 0% | 1,772 | 549 | -69% | 0 | 0 | — |
case-21 | fail→fail | 4,702 | 8,672 | +84% | 1 | 1 | 0% | 374 | 1,226 | +228% | 0 | 0 | — |
case-22 | pass→pass | 15,601 | 11,500 | -26% | 1 | 1 | 0% | 2,947 | 2,390 | -19% | 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 17 counted toward the lift figure. The other 5 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 +32 percentage points is the difference between those two pass rates over the 17 comparable cases. 2 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.