Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Regression-check citation anchoring (citations stay in the same subsection) to prevent “polish drift” that breaks claim→evidence alignment. **Trigger**: citation anchoring, citation drift, regression, cite stability, 引用锚定, 引用漂移. **Use when**: after editing/polishing, you want to confirm citations did not migrate across `###` subsections.
.claude/skills/willoscar-citation-anchoring/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-12 | ✗→✓ | ▲ Improved | -12% | 0% |
| case-06 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 181% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -39% | 0% |
Purpose: prevent a common failure mode: polishing rewrites text and accidentally moves citation markers into a different ### subsection, breaking claim→evidence alignment.
output/DRAFT.mdoutput/citation_anchors.prepolish.jsonl (baseline; created by draft-polisher on first run)output/CITATION_ANCHORING_REPORT.md (PASS/FAIL + drift examples)draft-polisher captures a baseline once per run: output/citation_anchors.prepolish.jsonl.Role:
Steps:
1) Load the baseline anchors. 2) Parse the current output/DRAFT.md into ### subsections and extract citation keys per subsection. 3) Compare current sets to baseline sets:
4) Write output/CITATION_ANCHORING_REPORT.md:
- Status: PASS only if no drift is detected- Status: FAIL with a short diff table + examplesIf you intentionally restructure across subsections:
output/citation_anchors.prepolish.jsonl and regenerate a new baseline (then treat that as the new regression anchor).Fix:
draft-polisher once to generate output/citation_anchors.prepolish.jsonl, then rerun the anchoring check.Fix:
output/citation_anchors.prepolish.jsonl and regenerate a new baseline (then treat that as the new regression anchor).| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-12 | fail→pass | 12,892 | 1,352 | -90% | 1 | 1 | 0% | 667 | 587 | -12% | 0 | 0 | — |
case-06 | fail→pass | 9,915 | 2,202 | -78% | 1 | 1 | 0% | 1,504 | 753 | -50% | 0 | 0 | — |
case-01 | fail→fail | 3,511 | 3,294 | -6% | 1 | 1 | 0% | 512 | 806 | +57% | 0 | 0 | — |
case-02 | fail→fail | 4,394 | 5,987 | +36% | 1 | 1 | 0% | 577 | 734 | +27% | 0 | 0 | — |
case-03 | fail→pass | 6,198 | 16,207 | +161% | 1 | 1 | 0% | 897 | 2,524 | +181% | 0 | 0 | — |
case-04 | fail→fail | 2,736 | 8,019 | +193% | 1 | 1 | 0% | 323 | 1,660 | +414% | 0 | 0 | — |
case-05 | fail→pass | 8,805 | 2,067 | -77% | 1 | 1 | 0% | 1,237 | 759 | -39% | 0 | 0 | — |
case-07 | pass→fail | 9,184 | 1,976 | -78% | 1 | 1 | 0% | 1,383 | 718 | -48% | 0 | 0 | — |
case-08 | fail→pass | 10,186 | 2,741 | -73% | 1 | 1 | 0% | 1,509 | 916 | -39% | 0 | 0 | — |
case-09 | fail→pass | 12,972 | 2,922 | -77% | 1 | 1 | 0% | 1,946 | 885 | -55% | 0 | 0 | — |
case-10 | fail→pass | 13,712 | 1,690 | -88% | 1 | 1 | 0% | 979 | 654 | -33% | 0 | 0 | — |
case-11 | fail→pass | 7,662 | 1,620 | -79% | 1 | 1 | 0% | 1,026 | 666 | -35% | 0 | 0 | — |
case-13 | fail→fail | 8,420 | 2,268 | -73% | 1 | 1 | 0% | 1,253 | 743 | -41% | 0 | 0 | — |
case-14 | fail→pass | 8,288 | 3,589 | -57% | 1 | 1 | 0% | 1,155 | 986 | -15% | 0 | 0 | — |
case-15 | fail→pass | 16,068 | 10,226 | -36% | 1 | 1 | 0% | 2,381 | 1,915 | -20% | 0 | 0 | — |
case-16 | pass→pass | 6,160 | 2,530 | -59% | 1 | 1 | 0% | 938 | 801 | -15% | 0 | 0 | — |
case-17 | pass→pass | 11,685 | 3,938 | -66% | 1 | 1 | 0% | 1,674 | 1,023 | -39% | 0 | 0 | — |
case-18 | fail→pass | 13,430 | 2,516 | -81% | 1 | 1 | 0% | 1,908 | 819 | -57% | 0 | 0 | — |
case-19 | fail→fail | 8,204 | 3,395 | -59% | 1 | 1 | 0% | 1,308 | 937 | -28% | 0 | 0 | — |
case-20 | fail→fail | 4,466 | 2,817 | -37% | 1 | 1 | 0% | 737 | 817 | +11% | 0 | 0 | — |
case-21 | fail→fail | 1,472 | 3,608 | +145% | 1 | 1 | 0% | 171 | 910 | +432% | 0 | 0 | — |
case-22 | pass→pass | 9,572 | 5,904 | -38% | 1 | 1 | 0% | 1,465 | 1,333 | -9% | 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 20 counted toward the lift figure. The other 2 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 +45 percentage points is the difference between those two pass rates over the 20 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.