Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Run a multi-stage extraction cascade (mention detection, document-level coreference clustering, optional saliency judgment, N-ary relation/triple extraction) directly over a paper's full text — covers SciERC, SciREX, and NLP Contribution Graph. Use this whenever cross-sentence or document-level entity/relation extraction is needed (e.g. SciREX-style Task-Dataset-Metric-Score tuples); do NOT use unit-classification for this, since these methods reason over the whole document's mentions, not indep
.claude/skills/yogsoth-ai-multi-stage-cascade-extraction/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -37% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -12% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -23% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -24% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 12% | 0% |
Mention detection → coreference clustering → saliency] → relation extraction, all stages consuming the full prior stage's output. Covers SciERC/SciREX/NLP-Contribution-Graph — three methods with different stage counts but the same "consume-the-full-prior-layer" structure (graph correction S6: merged under the unifying rule "same action-sequence length → mergeable via parameterization").
Subagent — spawned via spawn-agent skill.
This cascade discovers its own mention spans over the whole document rather than consuming pre-segmented sentence/clause units — sentence-level segmentation is the wrong granularity for a method whose relations are 99% cross-sentence (SciREX's own reported figure). This is a deliberate graph choice, not an oversight — see spec §5's flagged note before "fixing" this dependency.
NLP Contribution Graph's own reported consistency figures fall from stage to stage (67.92% → 41.82% → 22.31%) — this is the shared risk profile of this whole method family, not specific to one method. Producing every stage's intermediate output (not just the final relations) is what makes this compounding visible and debuggable.
<!-- BEGIN available-tables (generated) -->
| SOP | When to use | | --- | --- | | spawn-agent | Spawn a customized CC subagent with full MCP tool access. |
<!-- END available-tables (generated) -->
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 31,210 | 42,994 | +38% | 1 | 1 | 0% | 5,324 | 4,277 | -20% | 0 | 0 | — |
case-02 | fail→fail | 30,452 | 55,531 | +82% | 1 | 1 | 0% | 5,360 | 1,215 | -77% | 0 | 0 | — |
case-03 | fail→fail | 13,424 | 14,765 | +10% | 1 | 1 | 0% | 1,494 | 1,998 | +34% | 0 | 0 | — |
case-04 | pass→pass | 22,719 | 26,669 | +17% | 1 | 1 | 0% | 3,427 | 4,693 | +37% | 0 | 0 | — |
case-05 | pass→pass | 14,995 | 16,583 | +11% | 1 | 1 | 0% | 1,664 | 2,298 | +38% | 0 | 0 | — |
case-06 | pass→pass | 29,576 | 36,037 | +22% | 1 | 1 | 0% | 4,420 | 5,660 | +28% | 0 | 0 | — |
case-07 | fail→pass | 22,053 | 13,221 | -40% | 1 | 1 | 0% | 2,571 | 1,632 | -37% | 0 | 0 | — |
case-08 | fail→pass | 21,369 | 17,038 | -20% | 1 | 1 | 0% | 2,420 | 2,134 | -12% | 0 | 0 | — |
case-09 | fail→pass | 26,068 | 19,879 | -24% | 1 | 1 | 0% | 3,452 | 2,675 | -23% | 0 | 0 | — |
case-10 | fail→fail | 23,109 | 12,530 | -46% | 1 | 1 | 0% | 3,282 | 1,465 | -55% | 0 | 0 | — |
case-11 | fail→pass | 31,501 | 22,836 | -28% | 1 | 1 | 0% | 4,680 | 3,574 | -24% | 0 | 0 | — |
case-12 | fail→pass | 19,349 | 17,155 | -11% | 1 | 1 | 0% | 2,047 | 2,286 | +12% | 0 | 0 | — |
case-13 | fail→pass | 21,269 | 12,290 | -42% | 1 | 1 | 0% | 2,631 | 1,502 | -43% | 0 | 0 | — |
case-14 | fail→pass | 19,936 | 8,987 | -55% | 1 | 1 | 0% | 2,098 | 1,043 | -50% | 0 | 0 | — |
case-15 | pass→pass | 19,338 | 18,955 | -2% | 1 | 1 | 0% | 2,060 | 2,522 | +22% | 0 | 0 | — |
case-16 | pass→pass | 20,816 | 12,047 | -42% | 1 | 1 | 0% | 2,565 | 1,635 | -36% | 0 | 0 | — |
case-17 | fail→pass | 19,881 | 8,956 | -55% | 1 | 1 | 0% | 2,108 | 1,080 | -49% | 0 | 0 | — |
case-18 | pass→pass | 21,570 | 15,870 | -26% | 1 | 1 | 0% | 2,582 | 1,824 | -29% | 0 | 0 | — |
case-19 | fail→pass | 19,728 | 15,159 | -23% | 1 | 1 | 0% | 2,242 | 1,928 | -14% | 0 | 0 | — |
case-20 | fail→pass | 16,261 | 10,887 | -33% | 1 | 1 | 0% | 1,787 | 1,229 | -31% | 0 | 0 | — |
case-21 | pass→pass | 21,409 | 18,694 | -13% | 1 | 1 | 0% | 2,612 | 2,501 | -4% | 0 | 0 | — |
case-22 | fail→pass | 21,524 | 18,001 | -16% | 1 | 1 | 0% | 2,486 | 2,319 | -7% | 0 | 0 | — |
case-23 | fail→pass | 28,422 | 13,541 | -52% | 1 | 1 | 0% | 3,861 | 1,727 | -55% | 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 +52 percentage points is the difference between those two pass rates over the 23 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.