Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →LLM-based extraction of per-paper workflow graphs and content/structural reproducibility scoring.
.claude/skills/docxology-reproducibility-workflow-graph-assessment/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-21 | ✗→✓ | ▲ Improved | 9% | 0% |
| case-02 | ✗→✓ | ▲ Improved | 69% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -46% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -61% | 0% |
You are interacting with the src/reproducibility/ module. This decomposes a paper's own described pipeline (source/method/experiment/sink steps) into a workflow graph and scores how reproducible that pipeline is from the paper's own text.
ContentWeights, StructuralWeights) within Python scripts. Bind them to LLMConfig and the reproducibility_assessment block of config.yaml, loaded via config_loader.load_reproducibility_config().WorkflowNode must carry a non-empty source_quote copied verbatim from the paper's full text. Never accept or fabricate a node without one — this is the node's entire evidentiary basis.project_config.fulltext.enabled is false and no --fulltext-dir override is supplied, treat a reproducibility_assessment run as a no-op that still yields valid, empty-but-well-formed outputs — never mistake the warning for a crash.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-20 | fail→fail | 16,091 | 14,660 | -9% | 1 | 1 | 0% | 3,324 | 3,429 | +3% | 0 | 0 | — |
case-21 | fail→pass | 11,636 | 10,619 | -9% | 1 | 1 | 0% | 2,260 | 2,455 | +9% | 0 | 0 | — |
case-01 | fail→fail | 12,536 | 8,099 | -35% | 1 | 1 | 0% | 2,682 | 2,059 | -23% | 0 | 0 | — |
case-02 | fail→pass | 6,294 | 11,112 | +77% | 1 | 1 | 0% | 1,163 | 1,970 | +69% | 0 | 0 | — |
case-03 | fail→pass | 13,282 | 7,108 | -46% | 1 | 1 | 0% | 2,854 | 1,784 | -37% | 0 | 0 | — |
case-04 | fail→pass | 12,793 | 5,224 | -59% | 1 | 1 | 0% | 2,671 | 1,452 | -46% | 0 | 0 | — |
case-05 | fail→pass | 10,231 | 2,789 | -73% | 1 | 1 | 0% | 1,740 | 679 | -61% | 0 | 0 | — |
case-06 | fail→pass | 14,441 | 3,256 | -77% | 1 | 1 | 0% | 1,890 | 711 | -62% | 0 | 0 | — |
case-07 | pass→pass | 11,333 | 4,385 | -61% | 1 | 1 | 0% | 1,903 | 1,020 | -46% | 0 | 0 | — |
case-08 | pass→pass | 11,885 | 3,926 | -67% | 1 | 1 | 0% | 1,763 | 931 | -47% | 0 | 0 | — |
case-09 | fail→pass | 12,451 | 2,788 | -78% | 1 | 1 | 0% | 1,888 | 794 | -58% | 0 | 0 | — |
case-10 | fail→pass | 13,016 | 3,740 | -71% | 1 | 1 | 0% | 2,708 | 901 | -67% | 0 | 0 | — |
case-11 | fail→pass | 10,574 | 4,635 | -56% | 1 | 1 | 0% | 1,690 | 1,192 | -29% | 0 | 0 | — |
case-12 | fail→pass | 8,895 | 1,844 | -79% | 1 | 1 | 0% | 1,465 | 634 | -57% | 0 | 0 | — |
case-13 | fail→pass | 11,600 | 6,603 | -43% | 1 | 1 | 0% | 1,822 | 1,427 | -22% | 0 | 0 | — |
case-14 | fail→pass | 9,459 | 3,070 | -68% | 1 | 1 | 0% | 1,379 | 833 | -40% | 0 | 0 | — |
case-15 | pass→pass | 10,422 | 3,817 | -63% | 1 | 1 | 0% | 1,920 | 1,019 | -47% | 0 | 0 | — |
case-16 | fail→pass | 9,986 | 3,819 | -62% | 1 | 1 | 0% | 1,839 | 901 | -51% | 0 | 0 | — |
case-17 | fail→pass | 11,119 | 2,025 | -82% | 1 | 1 | 0% | 1,680 | 680 | -60% | 0 | 0 | — |
case-18 | fail→fail | 8,307 | 4,133 | -50% | 1 | 1 | 0% | 1,389 | 432 | -69% | 0 | 0 | — |
case-19 | fail→pass | 13,409 | 12,592 | -6% | 1 | 1 | 0% | 2,716 | 1,878 | -31% | 0 | 0 | — |
case-22 | fail→pass | 14,360 | 1,916 | -87% | 1 | 1 | 0% | 1,366 | 578 | -58% | 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 21 counted toward the lift figure. The other 1 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 +73 percentage points is the difference between those two pass rates over the 21 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.