Environmental intelligence, in production.

III.1

The streamflow model.

Seven-seed BF16 ensemble — an autoregressive nowcasting model, evaluated with full per-region distributions and honest reporting of where published baselines still lead.

Public benchmark: median Nash-Sutcliffe Efficiency 0.830 on a k-fold spatial hold-out — an autoregressive nowcasting setup that uses recent observed flow. On a held-out US CAMELS test period it reaches 0.800; on 181 held-out Canadian basins, 0.894. Across the full 16,299-basin Caravan assembly, 0.804.

Across the regional sub-sets we report the full per-region medNSE distribution rather than a single headline, worst-decile basins surfaced. We compare against published pure-simulation baselines — a different task from our autoregressive setup — and report where those baselines still lead us as honestly as where we close the gap, including CAMELS-GB, where a strong native-forcing regional baseline leads. No cherry-picking.

Built on public benchmark data — the Caravan v1.6 multi-source basin assembly — pinned at version, spanning 16,299 basins. Reproducibility details available to research collaborators on request.

III.3

Which pipe breaks next.

A calibrated risk score for every segment in a utility's network — validated across three independent cities on out-of-time data.

Across Calgary, Kitchener, and DC Water, the model ranks every pipe segment by how likely it is to fail next, validated on out-of-time data each city had never seen. AUROC lands between 0.875 and 0.910, and — crucially — the probabilities are calibrated (expected calibration error ≤ 0.008): a predicted 5% break-risk means roughly five in a hundred such segments actually break.

In practice it turns a utility's existing records into a ranked, replace-first list — so capital goes where the risk actually is, not where the last break happened. The same approach has held up across three independent cities.

→Methodology & calibration