Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use when auditing public research exemplars for deterministic publication readiness across tests, methods, evidence, artifacts, figures, manuscript sources, and generated outputs.
.claude/skills/docxology-publication-readiness-audit/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | -81% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -79% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 2% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -70% | 0% |
Run the shared audit after source, tests, methods, manuscript, or generated output changes. Deterministic failures block; review-required findings are explicit handoff items for a human editor or domain reviewer.
bashuv run python -m infrastructure.validation.cli publication-audit \ --all-public --strict --rendered --format markdown
The audit is read-only. Fix the producer or source contract, regenerate, and run it again. Source placeholders may remain when the canonical hydrated or combined rendered input is resolved. Do not edit reports or rendered provenance receipts by hand; stage 04 writes them after a green validation run, and scripts/maintenance/refresh_rendered_provenance.py backfills already-green tracked snapshots.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 15,470 | 4,786 | -69% | 1 | 1 | 0% | 2,630 | 392 | -85% | 0 | 0 | — |
case-02 | fail→fail | 7,823 | 4,917 | -37% | 1 | 1 | 0% | 1,268 | 312 | -75% | 0 | 0 | — |
case-03 | fail→pass | 10,204 | 3,052 | -70% | 1 | 1 | 0% | 1,828 | 351 | -81% | 0 | 0 | — |
case-04 | pass→pass | 8,362 | 2,805 | -66% | 1 | 1 | 0% | 1,412 | 685 | -51% | 0 | 0 | — |
case-05 | fail→pass | 8,436 | 2,059 | -76% | 1 | 1 | 0% | 1,163 | 527 | -55% | 0 | 0 | — |
case-06 | fail→pass | 14,247 | 1,657 | -88% | 1 | 1 | 0% | 2,071 | 425 | -79% | 0 | 0 | — |
case-07 | pass→pass | 11,327 | 3,944 | -65% | 1 | 1 | 0% | 1,641 | 778 | -53% | 0 | 0 | — |
case-08 | fail→pass | 19,079 | 4,226 | -78% | 1 | 1 | 0% | 950 | 965 | +2% | 0 | 0 | — |
case-09 | pass→pass | 7,579 | 2,285 | -70% | 1 | 1 | 0% | 1,275 | 481 | -62% | 0 | 0 | — |
case-10 | pass→pass | 12,775 | 1,820 | -86% | 1 | 1 | 0% | 1,973 | 441 | -78% | 0 | 0 | — |
case-11 | pass→pass | 16,006 | 1,584 | -90% | 1 | 1 | 0% | 2,364 | 434 | -82% | 0 | 0 | — |
case-12 | fail→pass | 8,635 | 1,815 | -79% | 1 | 1 | 0% | 1,525 | 452 | -70% | 0 | 0 | — |
case-13 | fail→pass | 13,820 | 1,586 | -89% | 1 | 1 | 0% | 1,053 | 427 | -59% | 0 | 0 | — |
case-14 | pass→pass | 12,315 | 8,207 | -33% | 1 | 1 | 0% | 1,914 | 1,514 | -21% | 0 | 0 | — |
case-15 | pass→pass | 5,320 | 2,835 | -47% | 1 | 1 | 0% | 882 | 718 | -19% | 0 | 0 | — |
case-16 | pass→pass | 8,338 | 3,993 | -52% | 1 | 1 | 0% | 1,222 | 726 | -41% | 0 | 0 | — |
case-17 | fail→pass | 11,627 | 5,888 | -49% | 1 | 1 | 0% | 1,971 | 1,063 | -46% | 0 | 0 | — |
case-18 | pass→pass | 9,545 | 3,935 | -59% | 1 | 1 | 0% | 1,587 | 824 | -48% | 0 | 0 | — |
case-19 | pass→pass | 5,633 | 5,304 | -6% | 1 | 1 | 0% | 964 | 1,058 | +10% | 0 | 0 | — |
case-20 | pass→pass | 5,973 | 4,057 | -32% | 1 | 1 | 0% | 955 | 887 | -7% | 0 | 0 | — |
case-21 | pass→pass | 3,683 | 3,100 | -16% | 1 | 1 | 0% | 587 | 623 | +6% | 0 | 0 | — |
case-22 | pass→pass | 3,697 | 3,919 | +6% | 1 | 1 | 0% | 574 | 801 | +40% | 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 19 counted toward the lift figure. The other 3 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 +32 percentage points is the difference between those two pass rates over the 19 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.