⚠️ In an affected area? Use official channels: National Weather Service National Hurricane Center FEMA Emergencies: 911 — Ray Resilience is a research tool and never replaces official warnings.
Ray Resilience · Resilience intelligence for every place · 面向每一个地点的灾害韧性智能 · OASIS @ ACM SIGSPATIAL 2026

Trust only what the evidence supports. Including about where it ends.

Ray Resilience is an AI-powered WebGIS and smartphone app. Its assistant, Ray, is a risk-analyst agent operating inside a verifiable harness that checks every spatial operation, preserves full provenance, and refuses claims beyond its evidence — with every unknown declared as prominently as every finding. Hazard monitoring is nationwide; exposure and damage analysis exist inside three deep-case areas and the app says so rather than extrapolating.

LIVE · hourly national watch (USGS · NWS · NHC · NIFC) + Day-1 flash-flood outlook (WPC) + three harness-checked deep cases: Eaton Fire · Hurricane Milton · Hurricane Ian
How It Works

Five stages, one chain of custody

Each stage consumes and produces auditable artifacts. Nothing is fabricated, nothing is skipped silently — the pipeline either delivers evidence or tells you exactly why it can't.

01 · LIVE
Watcher
Four keyless US connectors (USGS, NWS, NHC, NIFC/WFIGS) refresh hourly via CI — append-only gzip snapshots, per-source health, fail-closed on source errors.
LIVE · HOURLY
02 · PER EVENT
Dossier
Machine-readable event records with data sources, excluded sources (with reasons), and declared unknowns — attribution caveats stated per feature, not buried.
VERIFIED
03 · TIER 2
Exposure
Ground-truth structure damage (CAL FIRE DINS), county debris volumes, and CDC SVI joined onto H3 tiles — CRS asserted, join integrity checked, uncertainty mandatory.
VERIFIED
04 · TIER 3
Evidence
Reliability-gated cross-view damage grids from matched pre/post street-view imagery. Still fails closed: generated imagery is excluded and the exclusion is auditable.
VERIFIED
05 · DECISION
Decision
Damage × social-vulnerability trade-off sliders re-rank tile priorities client-side in the app; every human adjustment is audit-logged. Agent chat ships with the gateway.
IN THE APP
Deep Cases

Three events, two hazards, one harness

Every layer in the app traces back to hashed source snapshots through the Steward Harness — CRS asserted, joins verified, uncertainty mandatory, failures preserved in the audit log. Explore all three in the app. (Super Typhoon Bavi 2026, the project's first pre-event capture, is archived append-only and never deleted.)

3
Deep cases (wildfire + 2 hurricanes)
18,428
Ground-truth structure damage points (DINS)
5,686
Labeled cross-view evidence samples
134k
Source files SHA-256 hashed in the registry

Committed Case Records

Eaton Fire 2025 (CA)damage grid · SVI join · evidence
Hurricane Milton 2024 (FL)debris grid · dual-AOI evidence
Hurricane Ian 2022 (FL)evidence grid · sample density
Declared, not hiddenattribution caveats · excluded sources
Why Trust It

Honest by construction

Disaster tools earn trust by admitting what they don't know. Three rules are enforced in code, not just promised in prose.

Append-only provenance

Every artifact records its agent, UTC timestamp, and inputs. Pre-event products are frozen — post-event validation can't be quietly rewritten.

Fail closed

No imagery → no damage numbers. Missing inputs → a recorded failure with the reason. The pipeline never interpolates its way past missing evidence.

Declared unknowns

Every decision product lists what is not known as prominently as what is. An emergency manager deserves the full picture, including the gaps.

Who It Helps

Built to be useful, not just publishable

The same artifact chain serves three audiences without modification.

Emergency managers

Tile-level damage × social-vulnerability priorities with hands-on trade-off sliders, validity badges on every layer, and explicit unknowns. Inspection routes remain future work.

Researchers

A reproducible, fail-closed pipeline with tests and CI. The frozen pre-event record enables honest "did we predict the harm?" validation — rare in disaster ML.

  • Cross-view methodology: CrossViewGate research line
  • Entry: OASIS @ ACM SIGSPATIAL 2026, Track A

Affected communities

A hard rule: we point to official warning channels first. Ray Resilience augments official information; it never replaces it.

Get Started

Run the pipeline in two minutes

The test suite doubles as the verifiable evaluation environment: outcome checks, the policy matrix, and pipeline audits all run in CI on every push. Deep-case builders replay against the hashed dataset registry.

# install (h3 powers the tile grids; dev adds pytest)
pip install -e ".[deepcase,dev]"

# run the evaluation environment (363 Python tests; the app adds 67)
python -m pytest -q

# capture a live national watch snapshot (keyless)
python scripts/run_watch.py

# rebuild a deep case from the registered datasets
python scripts/build_eaton_case.py --data-root <registry root>

# run the app locally
cd app && npm install && npm run dev