Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when `evidence-review` has an extraction table and needs lightweight risk-of-bias fields. **Trigger**: bias, risk-of-bias, RoB, evidence quality, 偏倚评估, 证据质量. **Use when**: `evidence-review` 已生成 `papers/extraction_table.csv`，需要在 synthesis 前补齐偏倚/质量字段。 **Skip if**: 不是 evidence/systematic review，或还没有 `papers/extraction_table.csv`。 **Network**: none. **Guardrail**: 使用简单可复核刻度（low/unclear/high）+ 简短 notes；保持字段一致性。
.claude/skills/willoscar-bias-assessor/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-05 | ✗→✓ | ▲ Improved | -46% | 0% |
| case-14 | ✗→✓ | ▲ Improved | -60% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -38% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 44% | 0% |
Enriches the extraction table with lightweight risk-of-bias fields for evidence-review.
papers/extraction_table.csvpapers/extraction_table.csvThe table should expose stable RoB columns:
rob_selectionrob_measurementrob_confoundingrob_reportingrob_overallrob_notesAllowed values are fixed:
lowunclearhighscripts/run.py should:
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 3,439 | 3,556 | +3% | 1 | 1 | 0% | 167 | 401 | +140% | 0 | 0 | — |
case-02 | fail→fail | 4,463 | 4,375 | -2% | 1 | 1 | 0% | 193 | 405 | +110% | 0 | 0 | — |
case-03 | fail→fail | 16,610 | 4,631 | -72% | 1 | 1 | 0% | 1,862 | 369 | -80% | 0 | 0 | — |
case-04 | fail→fail | 8,259 | 3,471 | -58% | 1 | 1 | 0% | 1,423 | 682 | -52% | 0 | 0 | — |
case-05 | fail→pass | 12,424 | 5,309 | -57% | 1 | 1 | 0% | 2,036 | 1,107 | -46% | 0 | 0 | — |
case-14 | fail→pass | 8,985 | 2,295 | -74% | 1 | 1 | 0% | 1,384 | 558 | -60% | 0 | 0 | — |
case-06 | fail→pass | 15,700 | 6,749 | -57% | 1 | 1 | 0% | 2,429 | 1,523 | -37% | 0 | 0 | — |
case-07 | pass→pass | 9,231 | 3,152 | -66% | 1 | 1 | 0% | 1,446 | 675 | -53% | 0 | 0 | — |
case-08 | fail→fail | 4,817 | 2,562 | -47% | 1 | 1 | 0% | 674 | 576 | -15% | 0 | 0 | — |
case-09 | fail→fail | 10,543 | 3,296 | -69% | 1 | 1 | 0% | 1,550 | 660 | -57% | 0 | 0 | — |
case-20 | pass→fail | 12,158 | 6,492 | -47% | 1 | 1 | 0% | 1,671 | 1,265 | -24% | 0 | 0 | — |
case-10 | fail→pass | 10,717 | 4,724 | -56% | 1 | 1 | 0% | 1,601 | 991 | -38% | 0 | 0 | — |
case-11 | fail→fail | 6,961 | 3,842 | -45% | 1 | 1 | 0% | 1,034 | 782 | -24% | 0 | 0 | — |
case-12 | fail→pass | 10,412 | 16,343 | +57% | 1 | 1 | 0% | 1,645 | 2,371 | +44% | 0 | 0 | — |
case-13 | fail→pass | 7,586 | 2,911 | -62% | 1 | 1 | 0% | 1,210 | 677 | -44% | 0 | 0 | — |
case-15 | fail→pass | 9,762 | 2,495 | -74% | 1 | 1 | 0% | 1,557 | 593 | -62% | 0 | 0 | — |
case-16 | pass→pass | 11,070 | 2,649 | -76% | 1 | 1 | 0% | 1,830 | 628 | -66% | 0 | 0 | — |
case-17 | fail→pass | 9,612 | 2,500 | -74% | 1 | 1 | 0% | 1,547 | 624 | -60% | 0 | 0 | — |
case-18 | pass→pass | 5,213 | 1,500 | -71% | 1 | 1 | 0% | 770 | 389 | -49% | 0 | 0 | — |
case-19 | pass→pass | 9,257 | 2,563 | -72% | 1 | 1 | 0% | 1,394 | 631 | -55% | 0 | 0 | — |
case-21 | fail→fail | 6,346 | 7,050 | +11% | 1 | 1 | 0% | 312 | 427 | +37% | 0 | 0 | — |
case-22 | pass→fail | 21,304 | 5,767 | -73% | 1 | 1 | 0% | 4,049 | 434 | -89% | 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 +27 percentage points is the difference between those two pass rates over the 17 comparable cases. 4 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.