Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Samples context files, reads their L0 abstract (compressed ~100 token summary) and their full content, then evaluates whether the abstract accurately represents the content. Reports fidelity scores and flags abstracts that have drifted from their source.
.claude/skills/miosa-osa-verify/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-13 | ✗→✓ | ▲ Improved | -63% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -69% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -51% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 77% | 0% |
Cold-read L0 fidelity test — checks if abstracts accurately predict content.
Samples N contexts, takes title + L0 abstract only, scores match against full content. Dual path: LLM prediction scoring (Ollama) or Jaccard keyword similarity (fallback). Reports grades: A (≥0.8), B (≥0.6), C (≥0.4), D (≥0.2), F (<0.2).
/verify # Sample 20 contexts
/verify --sample 50 # Larger sample
/verify --node ai-masters # Specific node only| Flag | Description | Default | |------|-------------|---------| | --sample | Number of contexts to test | 20 | | --node | Limit to specific node | All |
bashcd engine && mix optimal.verify cd engine && mix optimal.verify --sample 50 --node ai-masters
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-13 | fail→pass | 12,019 | 3,781 | -69% | 1 | 1 | 0% | 1,977 | 728 | -63% | 0 | 0 | — |
case-23 | pass→pass | 21,381 | 12,938 | -39% | 1 | 1 | 0% | 2,326 | 2,338 | +1% | 0 | 0 | — |
case-07 | fail→pass | 7,882 | 2,106 | -73% | 1 | 1 | 0% | 1,329 | 592 | -55% | 0 | 0 | — |
case-01 | fail→pass | 9,780 | 1,957 | -80% | 1 | 1 | 0% | 1,574 | 495 | -69% | 0 | 0 | — |
case-02 | fail→pass | 11,058 | 3,850 | -65% | 1 | 1 | 0% | 1,608 | 785 | -51% | 0 | 0 | — |
case-03 | fail→pass | 3,745 | 3,766 | +1% | 1 | 1 | 0% | 478 | 845 | +77% | 0 | 0 | — |
case-04 | fail→pass | 17,000 | 2,041 | -88% | 1 | 1 | 0% | 2,764 | 439 | -84% | 0 | 0 | — |
case-05 | fail→pass | 22,016 | 2,402 | -89% | 1 | 1 | 0% | 1,525 | 595 | -61% | 0 | 0 | — |
case-06 | fail→pass | 7,246 | 2,084 | -71% | 1 | 1 | 0% | 979 | 568 | -42% | 0 | 0 | — |
case-08 | fail→pass | 7,868 | 2,393 | -70% | 1 | 1 | 0% | 1,186 | 584 | -51% | 0 | 0 | — |
case-09 | fail→pass | 9,174 | 3,700 | -60% | 1 | 1 | 0% | 1,406 | 671 | -52% | 0 | 0 | — |
case-10 | pass→pass | 6,495 | 1,973 | -70% | 1 | 1 | 0% | 726 | 583 | -20% | 0 | 0 | — |
case-11 | pass→pass | 12,608 | 1,810 | -86% | 1 | 1 | 0% | 2,162 | 503 | -77% | 0 | 0 | — |
case-12 | fail→pass | 12,352 | 1,732 | -86% | 1 | 1 | 0% | 1,776 | 492 | -72% | 0 | 0 | — |
case-14 | fail→pass | 20,340 | 7,518 | -63% | 1 | 1 | 0% | 1,706 | 640 | -62% | 0 | 0 | — |
case-15 | pass→pass | 9,899 | 2,253 | -77% | 1 | 1 | 0% | 1,439 | 586 | -59% | 0 | 0 | — |
case-16 | fail→pass | 9,816 | 1,779 | -82% | 1 | 1 | 0% | 1,395 | 494 | -65% | 0 | 0 | — |
case-17 | fail→pass | 12,717 | 1,908 | -85% | 1 | 1 | 0% | 1,815 | 505 | -72% | 0 | 0 | — |
case-18 | fail→pass | 5,999 | 1,870 | -69% | 1 | 1 | 0% | 872 | 503 | -42% | 0 | 0 | — |
case-19 | fail→pass | 7,681 | 1,833 | -76% | 1 | 1 | 0% | 1,092 | 526 | -52% | 0 | 0 | — |
case-20 | fail→pass | 11,237 | 2,642 | -76% | 1 | 1 | 0% | 1,839 | 542 | -71% | 0 | 0 | — |
case-21 | pass→pass | 14,009 | 5,793 | -59% | 1 | 1 | 0% | 2,526 | 1,158 | -54% | 0 | 0 | — |
case-22 | fail→fail | 14,500 | 6,764 | -53% | 1 | 1 | 0% | 2,587 | 1,478 | -43% | 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. The headline lift of +74 percentage points is the difference between those two pass rates over the 23 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.