Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use at study design when the supplied archive is about to become both the input and the thing you compare against. Covers where the comparison set has to come from, what a self-comparison cannot establish, and how to build a comparator from outside the shipped data.
.claude/skills/tangxiangru-earth-comparator-set-lives-outside-the-supplied-archive/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 29% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 4% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -27% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -10% | 0% |
| case-14 | ✗→✓ | ▲ Improved | 5% | 0% |
Budget acquisition time at design, before you write analysis code. In Earth science a number for a physical quantity is not a result until it stands beside independent estimates of the same quantity, and those never live in the folder you were handed. Plan to fetch four things: the other measurement techniques or instruments that estimate this quantity; the prior published assessments or synthesis values for it; the operational or reference baseline the field scores against (persistence, climatology, the incumbent product); and, where the claim concerns the future, a projection ensemble spanning the low and high forcing pathways the archive supports.
Plot every independent estimate's absolute value with its uncertainty interval, side by side, one panel per unit of analysis (region, basin, epoch, variable, lead). Only then add the pairwise differences. A plot of differences from your chosen reference is not an intercomparison, and it is the commonest way a correct set of numbers fails to register as one.
Express the same result in each normalisation the field uses, per-unit-area intensity, absolute total, and the impact-relevant equivalent, and state whether the headline is area- or mass-weighted. Repeat the entire analysis under every scenario the projection data supports, state the ordering across scenarios, report the share falling in each severity class, and carry it to the end-of-century horizon. Where your central estimate departs from a published one, lead with the ensemble range that brackets it and name the single factor carrying the gap.
All four Earth criteria the judge called absent (23.5% of the discipline) required a comparator that does not exist in data/: rival observation methods' absolute values, prior published assessments, competitor operational systems, and scenario projection ensembles. Agents treated the supplied folder as the study boundary; one fetched 100+ GB of external archives and spent it on its own novel ablation rather than the baseline comparison grid. It also addresses the Earth_002 mechanism where a central estimate departing from the published value scored 28 despite an ensemble range that contained it, because the abstract led with the point estimate.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 166,235 | 10,199 | -94% | 1 | 1 | 0% | 9,865 | 1,216 | -88% | 0 | 0 | — |
case-02 | fail→fail | 19,932 | 24,486 | +23% | 1 | 1 | 0% | 553 | 1,202 | +117% | 0 | 0 | — |
case-03 | fail→fail | 61,472 | 10,129 | -84% | 1 | 1 | 0% | 9,516 | 691 | -93% | 0 | 0 | — |
case-04 | pass→pass | 19,957 | 55,536 | +178% | 1 | 1 | 0% | 3,931 | 9,718 | +147% | 0 | 0 | — |
case-05 | pass→pass | 21,532 | 46,498 | +116% | 1 | 1 | 0% | 3,098 | 6,523 | +111% | 0 | 0 | — |
case-06 | pass→fail | 11,480 | 11,863 | +3% | 1 | 1 | 0% | 1,598 | 628 | -61% | 0 | 0 | — |
case-07 | fail→pass | 21,562 | 24,483 | +14% | 1 | 1 | 0% | 3,316 | 4,292 | +29% | 0 | 0 | — |
case-08 | fail→pass | 22,089 | 36,599 | +66% | 1 | 1 | 0% | 3,089 | 3,213 | +4% | 0 | 0 | — |
case-09 | pass→pass | 22,443 | 21,251 | -5% | 1 | 1 | 0% | 3,293 | 2,100 | -36% | 0 | 0 | — |
case-10 | fail→pass | 60,453 | 13,707 | -77% | 1 | 1 | 0% | 3,739 | 2,718 | -27% | 0 | 0 | — |
case-11 | fail→pass | 19,143 | 19,764 | +3% | 1 | 1 | 0% | 2,700 | 2,424 | -10% | 0 | 0 | — |
case-12 | pass→pass | 27,003 | 20,072 | -26% | 1 | 1 | 0% | 2,415 | 2,168 | -10% | 0 | 0 | — |
case-13 | fail→fail | 20,827 | 21,928 | +5% | 1 | 1 | 0% | 2,877 | 3,366 | +17% | 0 | 0 | — |
case-14 | fail→pass | 9,087 | 8,095 | -11% | 1 | 1 | 0% | 1,448 | 1,517 | +5% | 0 | 0 | — |
case-15 | fail→pass | 14,017 | 19,255 | +37% | 1 | 1 | 0% | 1,941 | 2,211 | +14% | 0 | 0 | — |
case-16 | fail→pass | 40,643 | 11,056 | -73% | 1 | 1 | 0% | 1,543 | 1,995 | +29% | 0 | 0 | — |
case-17 | pass→pass | 25,650 | 18,040 | -30% | 1 | 1 | 0% | 3,553 | 2,631 | -26% | 0 | 0 | — |
case-18 | fail→pass | 18,135 | 16,302 | -10% | 1 | 1 | 0% | 2,372 | 2,163 | -9% | 0 | 0 | — |
case-19 | pass→pass | 21,147 | 4,174 | -80% | 1 | 1 | 0% | 2,646 | 1,017 | -62% | 0 | 0 | — |
case-20 | pass→pass | 13,877 | 3,713 | -73% | 1 | 1 | 0% | 2,016 | 908 | -55% | 0 | 0 | — |
case-21 | fail→pass | 18,703 | 11,750 | -37% | 1 | 1 | 0% | 2,620 | 2,181 | -17% | 0 | 0 | — |
case-22 | fail→pass | 16,178 | 17,737 | +10% | 1 | 1 | 0% | 2,346 | 2,677 | +14% | 0 | 0 | — |
case-23 | pass→pass | 33,634 | 10,299 | -69% | 1 | 1 | 0% | 1,863 | 2,022 | +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. 23 cases were attempted, and 18 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 +39 percentage points is the difference between those two pass rates over the 18 comparable cases. 3 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.