Elysium Fields AIApplied AI · Environmental Risk
Get in touch
Field Almanac Platform Symposium Research About Get in touch
Vol. I — Field Almanac · 2026 49.51°N · 115.77°W · Cranbrook, BC

Reading the landscape, sharpening the question.

Elysium Fields AI is an applied AI research company based in Cranbrook, British Columbia. We build production systems for environmental risk.

§ 01What we build § 02How we work § 03Get in touch
Discipline
Applied AI research
Focus
Water, snow & infrastructure risk
Founded
2026 · Cranbrook, BC
Currently
Three systems live · one in beta
§ 01 — What we build Systems · in production

What we're building.

Three live prediction systems and a reasoning engine. Each is evaluated the same way — on test sets we lock before we start — and we report the misses next to the wins.

Pillar 01 · Environmental Live
I.

Environmental Intelligence

Production prediction systems for water, snow, and the infrastructure they put at risk.

HydroField predicts streamflow across a 16,299-basin training set drawn from public benchmark data (Caravan v1.6) — medNSE 0.830 on the public benchmark (an autoregressive nowcasting setup) and 0.894 on held-out Canadian basins, with full per-region distributions and the cases where published baselines still lead reported in full. AvalancheWatch posts a fresh regional danger rating every morning across 46 North American forecast zones, approaching the published Swiss operational accuracy range of 74–78%. Our newest system predicts which buried water mains will fail, validated across three North American utilities with calibrated probabilities. We document the methodology and lock the test sets before we start.

01 HydroField · public benchmark medNSE (k-fold spatial hold-out, autoregressive) 0.830
02 HydroField · medNSE on 181 held-out Canadian basins, zero training overlap 0.894
03 AvalancheWatch · North American forecast zones, updated daily 46
04 Water main-break · independent utilities validated, calibrated 3
→Read the platform notes
Supporting · Reasoning engine Beta · in active development
II.

Symposium

A multi-agent tournament that turns a hard question into a synthesized, citation-validated answer.

Nine specialist LLM personas argue across rounds; an adversary stress-tests every position. The orchestrator terminates on saturation rather than token budget. Retrievals come from arXiv and other public corpora; every claim in the final synthesis is traced back to its source and verified by a citation validator. We use Symposium internally to sharpen our own research decisions — and the same engine generalizes to any question that deserves more than one model's opinion.

01 Specialist personas per tournament 9
02 Retrieval corpora · arXiv plus expandable sources arXiv
03 Termination criterion · saturation, not token budget sat.
04 Citation validator on every synthesis claim cited
→Read how it works
§ 02 — How we work Three commitments

How we work.

For any number we report, we can hand a reviewer the whole reproduction chain — the data, the configuration, the git hash, the sacred test set, and the audit that closed the loop.

i. Pipeline consistency A single declared metric drives checkpoint selection, early stopping, learning-rate schedule, ensemble assembly, and the final reported number. No silent divergence between training and reporting.† † Internal Rule 21
Forensic-audit closed April 2026. Documentation available on request.
ii. Sacred test sets Every published number is computed once, on a test set defined before the first experiment in the series ran. No tuning on the holdout. The reproduction chain is preserved end-to-end.‡ ‡ HydroField sacred-test
626 basins · 6 Caravan datasets · locked April 2026. Anchor-test harness reproducible.
iii. Calibrated uncertainty Every model goes through a pre-publication calibration test: a 90% predicted interval should contain 90% of observed values, scored on the held-out test set. Models that pass ship as production claims; models that fail are reported as methodology findings, not buried.§ § Calibration
CRPS-scored. Pinball-quantile methodology, V2 cycle closed May 2026.
§ 03 — From the lab Recent

Field notes.

A working log. Milestones, methodology updates, and what we've recently locked, closed, or shipped. Quiet on the moat; substantive on the work.

2026 · Jun 8 Water main-break prediction validated across three North American utilities. Calibrated probabilities (ECE ≤ 0.008) on an out-of-time test. Benchmark
2026 · Jun 7 HydroField Canadian benchmark: medNSE 0.894 on 181 held-out basins, zero training overlap. Verifier-reproduced to 1e-6. Benchmark
2026 · May 24 HydroField public benchmark locked: medNSE 0.830 (k-fold spatial hold-out, autoregressive nowcasting). Published pure-sim baselines (Kratzert 2019, HydroDL) are a different task — we report where they still lead, including CAMELS-GB. Benchmark
2026 · Apr 18 Sacred test set locked for HydroField at 16,299 basins (Caravan v1.6). Anchor-test harness preserved. Discipline
§ 04 — Common questions FAQ

Questions, answered.

What does Elysium Fields AI do?
We build production machine-learning systems that predict environmental risk — streamflow and flood, avalanche danger, and water main-break — for the operators and planners who carry that risk. We are an applied AI research company in Cranbrook, British Columbia.
How accurate is HydroField, the streamflow model?
Public benchmark median NSE 0.830 on a k-fold spatial hold-out (an autoregressive nowcasting setup that uses recent observed flow), 0.894 on held-out Canadian basins, and 0.800 on a held-out US CAMELS test period. Published pure-simulation baselines (Kratzert 2019, HydroDL) address a different task; we report full per-region distributions and where those baselines still lead, including CAMELS-GB.
What is AvalancheWatch?
A live system that issues a daily avalanche-danger rating across 46 North American forecast zones. It reaches 73.2% exact five-level accuracy, approaching the published Swiss operational range of 74–78%, and 85.9% on binary “considerable-or-higher” alerts.
What does the water main-break model do?
It ranks which buried water mains are most likely to fail next, so a utility can replace pipe before it breaks. It is validated across three independent North American utilities with calibrated probabilities (expected calibration error ≤ 0.008; AUROC 0.875–0.910).
Who is Elysium Fields AI for?
Water utilities, ski and avalanche operations, flood and emergency managers, wildfire and drought planners, insurers, and researchers.
How do you keep the numbers trustworthy?
Every reported number is computed once on a test set locked before experimentation, one declared metric drives the whole training pipeline, and predicted uncertainty is calibrated — so results are reproducible and honestly reported, including where we don’t yet win.
§ 05 — Correspond

If our work is relevant to yours,
we'd like to hear from you.

Water utilities, snow and avalanche operations, flood and emergency managers, wildfire and drought planners, insurers, researchers, and fellow environmental scientists — whether you want an operational deployment where a system is already live, or to help shape what we build next, the door is open.

The systems How we work About