Do not draw one line.
The endpoint years contain different places, dates and exposure. First see the raw endpoint contrast. Then ask the published seasonal model a different question.
At 2016 the raw exposure-weighted endpoint means differ by 74.4 percent. That is not the published estimate. Switch to season adjusted to evaluate the 213-day projection.
Agreement is a choice of courtroom.
All three reduced contracts see the same 96 site-year summaries made from the 1,503 measured intervals and the same deterministic holdout. They disagree about whether time may remain after weather enters.
Season, habitat and a temporal trend
The common-row reduction gives season its curve, distinguishes the three published habitat clusters and lets year interact with season.
- weather windowsnone
- temporal structurelinear + seasonal interaction
- habitat matching2017 cluster
- validation20 held-out site-years
The 2025 Zenodo record was public but its files were restricted on 2026-08-25. The third tab restores a temporal term to the public 2024 common rows, one central rebuttal test. It does not impersonate the unavailable corrected habitat or NRW validation analysis.
The reduced 2024 weather contract has no autonomous time trend because year is excluded by construction. Its remaining residual slope is therefore evidence the reduction leaves temporal structure behind, not proof of a particular cause.
The check
Direct summation and a recurrence both produce a 76.700000 percent decline; delta below 1e-12.
The reconstruction lands on Hallmann et al.'s published 76.7 percent seasonal estimate within 0.05 percentage points.
Multiplying every 2016 prediction by 1.2 gives 72.04 percent, outside the anchor tolerance. The checker rejects it.
What can make this fail
- The two seasonal summation routes differ by at least 1e-12.
- The reconstructed decline misses 76.7 by more than 0.05 percentage points.
- The checker accepts the deliberately inflated 2016 curve.
- Any displayed court metric or source count drifts from the verifier's pinned summary.
Free choices and limits
- The visible curve is an anchor-calibrated Gaussian reconstruction, not a JAGS refit.
- The model court uses linear reductions and one every-fifth sorted site-year holdout.
- Malaise-trap biomass is not abundance, diversity, species loss or a global insect count.
- Cause and post-2016 direction remain disputed.
- Hallmann et al. 2017, article, S1/S2 data and model code.
- Müller et al. 2024, with public analysis bundle at Figshare 23536551.
- Hallmann et al. 2025, with restricted-file Zenodo record 14258659.