Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Provisions the oracle ML inference daemon with onnxruntime via uv. Use when setting up local ONNX model inference for skill quality evaluation.
.claude/skills/athola-setup/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | 25% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -63% | 0% |
| case-13 | ✗→✓ | ▲ Improved | -68% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-15 | ✗→✓ | ▲ Improved | -60% | 0% |
Provision the ML inference environment.
abstract:skills-eval)bashcd plugins/oracle && uv run python -c " from oracle.provision import provision_venv, get_venv_path result = provision_venv(get_venv_path()) print(result.message) "
provision_venv() returns a result with a non-error messageand the venv path exists on disk under plugins/oracle/
onnxruntime is importable inside the provisioned venv;verified by uv run python -c "import onnxruntime" exiting 0
start on next session, or a clear failure message citing the specific error and whether it is a uv or network issue
prevent a clean retry
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | pass→pass | 17,609 | 14,708 | -16% | 1 | 1 | 0% | 3,734 | 3,495 | -6% | 0 | 0 | — |
case-01 | fail→fail | 7,101 | 5,393 | -24% | 1 | 1 | 0% | 1,291 | 653 | -49% | 0 | 0 | — |
case-02 | fail→fail | 9,349 | 4,830 | -48% | 1 | 1 | 0% | 1,596 | 660 | -59% | 0 | 0 | — |
case-03 | fail→fail | 10,478 | 5,336 | -49% | 1 | 1 | 0% | 1,782 | 706 | -60% | 0 | 0 | — |
case-04 | pass→pass | 6,675 | 5,060 | -24% | 1 | 1 | 0% | 1,082 | 1,136 | +5% | 0 | 0 | — |
case-05 | fail→pass | 17,057 | 16,288 | -5% | 1 | 1 | 0% | 2,791 | 3,480 | +25% | 0 | 0 | — |
case-07 | fail→fail | 5,566 | 6,749 | +21% | 1 | 1 | 0% | 965 | 780 | -19% | 0 | 0 | — |
case-08 | fail→fail | 10,683 | 7,757 | -27% | 1 | 1 | 0% | 1,825 | 734 | -60% | 0 | 0 | — |
case-09 | fail→fail | 12,877 | 6,688 | -48% | 1 | 1 | 0% | 2,541 | 698 | -73% | 0 | 0 | — |
case-10 | fail→pass | 14,426 | 4,507 | -69% | 1 | 1 | 0% | 2,355 | 875 | -63% | 0 | 0 | — |
case-11 | fail→fail | 15,819 | 5,045 | -68% | 1 | 1 | 0% | 1,149 | 574 | -50% | 0 | 0 | — |
case-12 | fail→fail | 20,118 | 4,269 | -79% | 1 | 1 | 0% | 3,257 | 488 | -85% | 0 | 0 | — |
case-13 | fail→pass | 10,672 | 1,434 | -87% | 1 | 1 | 0% | 1,755 | 553 | -68% | 0 | 0 | — |
case-14 | fail→pass | 5,327 | 4,334 | -19% | 1 | 1 | 0% | 948 | 1,163 | +23% | 0 | 0 | — |
case-15 | fail→pass | 11,257 | 2,760 | -75% | 1 | 1 | 0% | 1,783 | 720 | -60% | 0 | 0 | — |
case-16 | fail→pass | 12,571 | 1,729 | -86% | 1 | 1 | 0% | 2,054 | 545 | -73% | 0 | 0 | — |
case-17 | fail→pass | 5,056 | 1,523 | -70% | 1 | 1 | 0% | 852 | 575 | -33% | 0 | 0 | — |
case-18 | fail→fail | 14,609 | 4,023 | -72% | 1 | 1 | 0% | 2,459 | 850 | -65% | 0 | 0 | — |
case-19 | pass→pass | 13,426 | 5,272 | -61% | 1 | 1 | 0% | 2,027 | 1,112 | -45% | 0 | 0 | — |
case-20 | fail→pass | 8,466 | 2,071 | -76% | 1 | 1 | 0% | 1,319 | 626 | -53% | 0 | 0 | — |
case-21 | pass→pass | 10,466 | 1,314 | -87% | 1 | 1 | 0% | 1,620 | 524 | -68% | 0 | 0 | — |
case-22 | fail→pass | 8,726 | 2,836 | -67% | 1 | 1 | 0% | 1,235 | 775 | -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 14 counted toward the lift figure. The other 8 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 +41 percentage points is the difference between those two pass rates over the 14 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.