Geosphere
Whether the corridor is releasing more carbon than is normal for the season. The CaCO₃ research quantifies how much of a source could be pulled back.
Inputs in-distribution, members in agreement, interval tight — the Trust Core stands behind this number.
Stress basis: NEE anomaly vs the same season in prior years: -1.99 gC·m⁻²·d⁻¹ versus the 8-year seasonal mean of -0.34.
Ensemble spread
Six heterogeneous members — a GRU, an MLP, a ridge autoregression, Holt smoothing, a seasonal baseline and persistence — each vote on the next value. Their weighted scatter relative to the recent one-step error is the disagreement signal; wide scatter means the members have genuinely split.
Trust signals
Spread of members vs expected one-step error.
Standardised distance of today's 11-feature input row from the training set. Furthest feature: Recent history of the target (-3.21σ).
The combined trust the Core assigns this output.
What drove this prediction
Input×Gradient attribution: the gradient of the neural members' forecast with respect to every input, multiplied by the input, summed over the window — so each named driver carries a sign and a magnitude. The AR member's coefficients fold into the target's own history.
The series' own recent values (autoregressive memory)
Light-use-efficiency GPP
Day-of-year phase
MODIS NDVI interpolated to the day
ERA5 daily shortwave radiation
ERA5 daily max VPD, the drought-stress driver
Monthly rain, the soil-moisture proxy
ERA5 daily mean 2 m temperature
Q10 respiration term
Day-of-year phase
ERA5 daily minimum, the MOD17 cold-limit driver
This stream, audited
7-step path
The ensemble is rolled forward recursively; each horizon carries its own conformal quantile from the calibration window, scaled by the adaptive factor, so the band widens honestly with lead time. Drivers are held at their last value and only the season advances.
Rolling backtest, last 90 steps
Every point is a real one-step forecast made from the data available at the time, against what the feed then reported. Misses outside the band are marked. Adaptive conformal inference nudges α each step toward the 90 % target — currently α=0.031 (band ×1.48 vs static).
Model card
trained 2026-10-02A heterogeneous ensemble: two neural networks trained by backpropagation on the archive, two classical forecasters, and two baselines every member must beat. Weights are set by validation error, never by the test window.
| Member | Kind | Weight | Test MAE* | Vote |
|---|---|---|---|---|
| GRU recurrent net1,941 params | neural · recurrent | 23% | 0.586 | -2.36 |
| MLP feed-forward net1,869 params | neural · feed-forward | 22% | 0.591 | -2.75 |
| AR(7) ridge | statistical | 19% | 0.619 | -2.10 |
| Holt smoothing | statistical | 16% | 0.645 | -2.75 |
| Seasonal naive | baseline | 8% | 1.101 | 0.57 |
| Persistence | baseline | 12% | 0.605 | -3.26 |
* held-out test window, native units
Carbon balance
The light-use-efficiency model's decomposition of today's flux, the corridor's trailing-year balance, and the global CO₂ background from Mauna Loa.
Provenance
Governed quantity: Next-day corridor-mean net ecosystem exchange (positive = source to the atmosphere).
archive pulled 2026-10-02 09:54 UTC · live tail merged at request time · all sources publicly available · nothing on this page is simulated