The research layer behind AI-native water intelligence.
Impact Research supports the product and trust surfaces of ALCUB3 Impact — Water Pulse, Observatory, BasinKit, and the public trust layer. One track. Four sections — methodology, benchmarks, reports, sources — each with its own discipline.
How water intelligence is measured and checked.
Methods is the public surface for Impact's scoring logic, evidence quality, uncertainty handling, and publication standards. Every claim on the Impact surface should trace back to something on this page.
Composite scoring
45% water quality, 30% drought risk, 25% flood risk. Each component carries its own tier (measured/estimated/modeled) and its own last-updated timestamp. The weights are versioned — change them and the version bumps.
AI water footprint
Stress-weighted AI footprint methodology. Raw liters are multiplied by local water stress so a query in drought-stressed Phoenix doesn't read the same as the same query in wet-year Iowa.
Satellite-derived change
Satellite water-body segmentation, forecast confidence, and regional applicability. Modeled outputs with explicit error ranges — trustable in aggregate, handled carefully per-instance.
Evidence tiers (full definitions)
- Measured: the claim comes from a real observation. EPA violations, USGS streamflow readings, lab test results. Strongest tier; latency is the main caveat.
- Estimated: reasonable math applied to observed inputs, but the specific case hasn't been verified. Population-weighted drought severity is estimated. Most good engineering.
- Modeled: a machine learning model or statistical projection produced the output. Satellite segmentation, PFAS risk projections, climate-adjusted forecasts. Useful in aggregate, careful per-instance.
- Roadmap: we intend to build this. It doesn't exist today. Transparent about what's planned vs shipped.
Water intelligence accuracy.
Impact benchmarks measure what the water-intelligence layer can actually claim. Coverage, calibration, confidence, regional generalization, and evidence-tier correctness.
Where the data reaches
Percentage of US zip codes covered by EPA-reported water systems, USGS gauge proximity, NOAA flood-alert coverage, and drought-monitor resolution. Where the data reaches, and where it doesn't.
Score vs ground truth
How the Water Health Score correlates with verified quality events — validated violations, confirmed contamination episodes, and downstream health outcomes where data exists.
Does the method travel?
How well methodologies developed on one region (e.g., California) generalize to another (e.g., Gulf Coast). Regional caveats are published with every score.
The water-intelligence landscape.
Public synthesis reports covering the water-intelligence field — methods, competitor framing, PFAS, groundwater, flood, infrastructure, and observability signals. The layer that feeds Water Pulse, BasinKit, and the Water Intelligence API.
Water intelligence research landscape 2026
Methods, competitor framing, PFAS, groundwater, flood, infrastructure, and observability signals. 100+ papers synthesized into 40 curated sources with five research hypotheses.
The three gaps that shape the product
Coverage gap (private wells, small systems), latency gap (how old is the data), interpretation gap (unreadable for non-specialists). These gaps define what Impact exists to close.
Global AI water consumption
Training GPT-3 ~700K liters. Projected 4.2–6.6 billion cubic meters per year by 2027 — more than Denmark. Translating those figures into a personal-level footprint with clear assumptions and visible caveats.
Environmental stack synthesis
What to use, what to reference, and how NOT to let infrastructure become the brand story. Informs Impact's use of NVIDIA OSS without NVIDIA becoming the identity.
Public data and source notes.
Every public surface traces back to these primary sources. Impact does not invent data — it integrates and interprets public sources honestly, with each source's coverage and limits named explicitly.
Drinking water data
EPA ECHO, Safe Drinking Water Information System (SDWIS), and violation records. Strongest for public water systems serving 25+ people. Does not cover private wells.
Surface water + streamflow
USGS gauges, daily values, and the National Water Information System (NWIS). Streamflow and water levels at gauge stations. Surface water only — not a groundwater surface.
Drought + flood alerts
US Drought Monitor (weekly), National Water Prediction Service, NOAA flood-alert feeds. Near real-time flood; weekly drought. Both labeled by last-updated timestamp.
Global water risk + observation
World Resources Institute Aqueduct, Sentinel-2 + Landsat imagery, and research-backed satellite segmentation models. Powers the Observatory signals layer.
Public provenance rules
- Source families: canonical datasets, papers, and external references each method depends on.
- Scope boundaries: what a source can support directly and where assumptions begin.
- Refresh expectations: whether a source is static, periodically updated, or actively monitored.
- Disclosure: if we can't trace a claim to a source, we don't publish the claim.
Where impact research lands.
See the product
Impact Research directly supports Water Pulse, Observatory, BasinKit, and the Water Intelligence API. The methodology and benchmarks on this page are what let those surfaces claim what they claim.
Open ALCUB3 Impact →Learn the vocabulary
The Water Intelligence with Impact path teaches readers how to interpret the evidence tiers on a real consumer product. Research explains why; the Institute teaches how.
Explore Impact →See the other track
Platform Research mirrors this structure for the core ALCUB3 products. Same methodology discipline, different domain.
Open Platform Track →