Resilience intelligence for every place. An accountable GeoAI system for place-based disaster intelligence — a WebGIS / smartphone app with an assistant, Ray, that may only say what the evidence supports, and a Steward Harness that enforces it in code.
Live site · Open the app · For judges: JUDGES.md · Paper · Release v0.1.0-oasis
▶ Watch the 2:55 demo — resident mode, the planner's trade-off, lineage, and Ask Ray, recorded live with voice-over and subtitles (how it was made).
Most GeoAI disaster systems demonstrate autonomy: a language model that runs a spatial pipeline. Ray Resilience demonstrates accountability. Every map layer traces back to a hashed source snapshot; every sentence the assistant produces cites an artifact or is refused; a declarative policy decides who may be told what, at which resolution; and a CI gate keeps the public build inside that policy. When the system does not know, it says so — as prominently as when it does.
 |
 |
| Resident mode. An address gets one of three answers — covered, outside the evaluated areas, or not determined — and, when covered, a plain-language dossier for its tile: residents, structures assessed, destroyed share, social vulnerability, critical facilities within 1 km, and the declared unknowns. Never a statement about a single parcel. | Planner mode. Damage × social-vulnerability priority on an H3 grid. The trade-off t · damage + (1 − t) · SVI is the planner's to set on a slider; every move is audit-logged. Lineage runs from any layer back to hashed source snapshots, including a rejected run the harness kept rather than erased. |
- Nationwide hourly watch — USGS earthquakes, NWS alerts, NHC tropical cyclones, NIFC/WFIGS wildfires, plus the NOAA/WPC Day-1 excessive-rainfall outlook as a separate, clearly bounded product. Per-source failures are declared, not hidden.
- Three deep cases, one harness — the Eaton Fire (2025, CA), Hurricane Milton (2024, FL) and Hurricane Ian (2022, FL): structure-damage ground truth, county debris volumes, CDC social vulnerability, 2020 census population and reliability-gated cross-view street imagery on H3 r9 grids. Analysis exists inside these areas and nowhere else, and the app says so.
- Ask Ray — a risk-analyst agent behind two deterministic gates: a policy pre-check before any model call, and a citation post-check after it. Refusals name the rule that triggered them. Works with any OpenAI-compatible endpoint; a local open model over Ollama by default, no key required.
- Model output as governed evidence — six open vision–language models graded 1,225 labelled street-view samples each under RAPID's verbatim prompts. The results are published as an evaluation and refused as a claim:
docs/vlm_model_comparison.md. - The harness caught its own system three times — a parcel-level file on the public site, an uncited reassurance that passed, an evidence store that read what the publisher withheld. Each incident is preserved and each fix is a control, not a patch:
docs/incidents/,docs/STATUS.md. - A verifiable evaluation environment — 363 Python and 67 app tests run in CI on every push, including a cell-by-cell policy-matrix sweep, adversarial claim checks, connector tests against real error envelopes, and an allowlist drift check.
Three loosely coupled planes over one harness. The data plane (GitHub Actions) runs the connectors and the deep-case builders and publishes hashed artifacts. The presentation plane is an installable, keyless, offline-capable PWA (React + Vite + MapLibre GL). The agent plane wraps any OpenAI-compatible model in policy pre-check → grounded generation → citation post-check → audit. If the agent is down, the maps keep working: graceful degradation is fail-closed design made visible.
The Steward Harness enforces four things between every stage and around every sentence:
| What is enforced | |
|---|---|
| Outcome validity | Executable spatial checks — CRS assertions, join integrity, sanity bounds, and a mandatory uncertainty block on every tile (a cell without one fails the build). |
| Process validity | SHA-256 lineage in an append-only audit log; failures recorded, never erased; the gateway's audit stores a point as its H3 cell and a question as a digest. |
| Institutional validity | One YAML policy, two planes of ordered rules with default deny. The claim plane scopes what may be asserted by role, evidence tier, resolution and geography; the distribution plane scopes what a build may publish, and CI fails the deploy on violation. |
| Verifiability | Each artifact carries retained > re-derivable > cited-only (weakest link) and a license attribute, so a claim resting on evidence nobody may keep says so. |
The policy is one file: src/geosteward/harness/policy_v1.yaml.
The full mechanism is documented in the technical manual.
| Case | Committed products | Declared limits |
|---|---|---|
| Eaton Fire 2025 (CA, wildfire) | 18,428 CAL FIRE DINS points → 265-cell damage grid · CDC SVI 2022 across 20 tracts · 2,244 cross-view samples → 109-cell evidence grid · 46,341 residents · 27 OSM facilities | 40 inaccessible points; the repairable class has n = 30; tract-to-cell SVI is a declared downscaling |
| Hurricane Milton 2024 (FL) | 2,556 labelled pre/post street-view pairs → 15-cell grid · 5,618 Pinellas cells with county debris volumes · 772,293 residents · 200 OSM facilities | post-event imagery is season-cumulative (Debby, Helene, Milton); a generated-imagery set was excluded, auditably |
| Hurricane Ian 2022 (FL) | 886 matched samples → 190-cell evidence grid · 4,121 street-view positions as a density-only layer · 5,428 residents · 79 OSM facilities | density cells support coverage, not point-level severity |
All source imagery traces to a hashed dataset registry (134,272 files, ~33 GB, SHA-256). Facilities are OpenStreetMap presence, never operational status. Population outside the evaluated tiles is reported as a total, never mapped into cells the event has no evidence for.
git clone https://github.com/rayford295/ray-resilience && cd ray-resilience
python -m pip install -e ".[deepcase,dev]"
python -m pytest -q # the evaluation environment: 363 tests (CI runs the same suite with unittest)
python scripts/publication_boundary.py plan # which artifacts may be published, and why
cd app && npm ci && npm test && npm run dev # the PWA at http://localhost:5173The app serves the committed artifacts — no keys, no services. To talk to Ray as well:
ollama pull gpt-oss:20b # or any OpenAI-compatible endpoint
python -m pip install -e ".[deepcase,gateway]"
uvicorn gateway.main:app --port 8080The deep-case builders read a ~33 GB corpus that is not redistributed here; third parties can verify every committed artifact, hash and audit row, and can re-run the VLM evaluations from the committed prediction records without a GPU.
├── app/ # PWA: resident + planner modes, lineage viewer, Ask Ray panel
├── gateway/ # FastAPI agent gateway — LLM-agnostic, harness middleware
├── src/geosteward/ # pipeline · connectors · deep-case builders · harness/ (policy, audit, publication)
├── events/ # eaton-2025/ · milton-2024/ · ian-2022/ · palisades-2025/ (evaluation) · archive/
├── docs/ # STATUS.md · manual/ · incidents/ · design/ · demo/ · vlm_model_comparison.md
├── paper/ # OASIS Track A paper (LaTeX + Word draft), figures, eligibility statement
├── scripts/ # builders, the publication boundary, the VLM sweep and comparison
└── tests/ # 363 tests — doubles as the verifiable evaluation environment
- RAY: Resilience Assistant for You — An Accountable GeoAI System for Place-Based Disaster Intelligence, OASIS Challenge @ ACM SIGSPATIAL 2026, Track A:
paper/ray-resilience-oasis2026.pdf(source and co-author Word draft alongside). - Eligibility and contribution statement:
paper/eligibility-and-contribution-statement.pdf. - Demo video (2:55, voice-over + subtitles):
docs/demo/ray-resilience-demo.mp4— a scripted, genuine walk-through produced bydocs/demo/record_demo.py; how it is made is indocs/demo/README.md. - Six-model VLM comparison, regenerated from the committed evaluation files:
docs/vlm_model_comparison.md.
Built on the RAPID line (Yang et al., 2026) — its prompts, metric and acceptance rules become harness stages; its LLM task planner is deliberately not adopted.
Yifan Yang (Geography; team lead and corresponding author) · Ziyi Wang (Computer Science and Engineering) · Wenjing Gong (Landscape Architecture and Urban Planning) · Lei Zou (Geography; faculty advisor) — Texas A&M University.
This research is supported by the National Academies of Sciences, Engineering, and Medicine
Gulf Research Program (SCON-10000653, SCON-10001536) and the U.S. National Science Foundation
(2318206). Claude (Anthropic) was used as a coding and writing assistant; commits it contributed
to carry a Co-Authored-By trailer, and the authors take full responsibility for the code and text.
MIT License. Ray Resilience is a research prototype — not an official forecasting or warning service. In an emergency, follow official guidance (National Weather Service, National Hurricane Center, FEMA, and local emergency management).