Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Design and maintain GRACE 4 verification entries, commands, scenarios, markers, and assertion evidence under .grace/verification.
.claude/skills/osovv-grace-verification/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 8% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -19% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -47% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -20% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -34% | 0% |
<skill> <purpose> Strengthen deterministic verification for modules and changes. Verification state lives in .grace/verification/index.xml and routed verification documents. Each durable module should have deterministic V-M-* coverage unless an explicit exception is planned. </purpose>
<workflow>
.grace/graph anchors and current V-M-* entries..grace/verification changes through the active change plan.</workflow> <cwd_contract> When verification commands run from a workspace or package directory, add one direct <Cwd>relative/project/path</Cwd> child to the owning V-M-* entry. Keep declared <TestFiles><File>...</File></TestFiles> paths project-root-relative; the CLI uses Cwd only to compare them with cwd-relative command arguments. </cwd_contract> <evidence_contract> Use <Marker> when module health must prove a runtime log or trace emission from linked implementation code. Use <TraceAssertion> for deterministic test or trace evidence that does not require runtime logging, such as pure functions, type-level modules, and core libraries. A non-empty marker or trace assertion satisfies the module-health evidence requirement; only authored markers require matching runtime emission and BLOCK_* evidence. </evidence_contract> </skill>
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 17,661 | 14,081 | -20% | 1 | 1 | 0% | 3,465 | 548 | -84% | 0 | 0 | — |
case-02 | fail→fail | 19,739 | 4,426 | -78% | 1 | 1 | 0% | 3,932 | 586 | -85% | 0 | 0 | — |
case-03 | fail→pass | 23,177 | 4,734 | -80% | 1 | 1 | 0% | 1,171 | 1,266 | +8% | 0 | 0 | — |
case-04 | pass→fail | 10,758 | 4,558 | -58% | 1 | 1 | 0% | 2,457 | 585 | -76% | 0 | 0 | — |
case-05 | pass→pass | 7,723 | 10,254 | +33% | 1 | 1 | 0% | 1,511 | 2,269 | +50% | 0 | 0 | — |
case-06 | pass→pass | 4,187 | 7,285 | +74% | 1 | 1 | 0% | 765 | 1,264 | +65% | 0 | 0 | — |
case-07 | fail→pass | 8,908 | 4,488 | -50% | 1 | 1 | 0% | 1,637 | 1,331 | -19% | 0 | 0 | — |
case-08 | pass→pass | 9,102 | 2,350 | -74% | 1 | 1 | 0% | 1,480 | 827 | -44% | 0 | 0 | — |
case-09 | fail→pass | 8,771 | 2,450 | -72% | 1 | 1 | 0% | 1,516 | 801 | -47% | 0 | 0 | — |
case-10 | fail→pass | 9,137 | 5,110 | -44% | 1 | 1 | 0% | 1,642 | 1,317 | -20% | 0 | 0 | — |
case-11 | fail→pass | 9,769 | 3,867 | -60% | 1 | 1 | 0% | 1,631 | 1,081 | -34% | 0 | 0 | — |
case-12 | fail→pass | 5,121 | 1,536 | -70% | 1 | 1 | 0% | 837 | 636 | -24% | 0 | 0 | — |
case-13 | fail→pass | 9,596 | 1,123 | -88% | 1 | 1 | 0% | 1,767 | 504 | -71% | 0 | 0 | — |
case-14 | pass→pass | 12,424 | 4,658 | -63% | 1 | 1 | 0% | 1,884 | 1,116 | -41% | 0 | 0 | — |
case-15 | pass→pass | 8,874 | 2,995 | -66% | 1 | 1 | 0% | 1,547 | 918 | -41% | 0 | 0 | — |
case-16 | fail→pass | 22,332 | 4,896 | -78% | 1 | 1 | 0% | 1,200 | 701 | -42% | 0 | 0 | — |
case-17 | fail→pass | 7,962 | 3,627 | -54% | 1 | 1 | 0% | 1,140 | 916 | -20% | 0 | 0 | — |
case-18 | pass→pass | 8,974 | 2,899 | -68% | 1 | 1 | 0% | 1,449 | 863 | -40% | 0 | 0 | — |
case-19 | fail→pass | 10,017 | 3,768 | -62% | 1 | 1 | 0% | 1,718 | 1,012 | -41% | 0 | 0 | — |
case-20 | fail→pass | 6,286 | 3,686 | -41% | 1 | 1 | 0% | 1,231 | 1,070 | -13% | 0 | 0 | — |
case-21 | fail→pass | 9,937 | 1,992 | -80% | 1 | 1 | 0% | 1,535 | 684 | -55% | 0 | 0 | — |
case-22 | fail→pass | 8,493 | 2,684 | -68% | 1 | 1 | 0% | 1,274 | 797 | -37% | 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 +55 percentage points is the difference between those two pass rates over the 17 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.