plateau-triage 0.1.4

Post-disaster building damage record layer: per-building damage state (danger / caution / inspected / unsurveyed) + confidence + provenance, bound to persistent PLATEAU building IDs. Honest by construction: covered=false ⇒ value=null.

plateau-triage 🚩

English · 日本語

After an earthquake, which buildings do you send rescuers to first —
and which ones has nobody even checked yet?

What it is

After a disaster a city has tens of thousands of buildings and far too few inspectors. plateau-triage turns disaster imagery into a building-by-building map of damage — and its one rule is that a building nobody has checked is never allowed to look "fine".

The trap it fixes: most damage maps leave un-inspected buildings blank — which looks exactly like "safe", so responders skip them. plateau-triage instead gives every building one of four explicit labels:

🔴 danger · 🟡 caution · 🟢 inspected — safe · ⬜ not yet surveyed

The grey "not yet surveyed" is the whole point: a building no one has looked at is shown as unknown — never assumed safe. A coverage meter tells you how much of the city is still grey ("8,000 / 30,000 checked — 73% unknown"), so the question becomes "what haven't we checked?" instead of "no red marks, must be fine."

Click any building to see where its status came from

Every label is traceable: which image, when, how confident, and the official building ID it's bound to. A grey building says, in plain words, "nobody has looked here — not assumed safe."

Who uses it, and how

1. A response HQ opens the map and reads the coverage meter — danger first, and "not yet surveyed" ranked above "confirmed safe", because an unknown building may be the most dangerous one on the block.
2. An inspector clicks a building to see where its status came from, decides whether to trust it or go look, and the record updates after the site visit.
3. A relief desk exports the records straight into existing systems — a screening feed for the victim-support system (被災者支援システム) and a pre-fill for the disaster-victim certificate (罹災証明). Unsurveyed buildings carry no claim, so the system can never issue a certificate for a building nobody checked.

Data comes either from public satellite datasets (bulk import) or from drone / street capture in the field — and when the drone can't see a building clearly, that building is honestly marked not yet surveyed, not guessed.

Quickstart

pip install plateau-triage          # from PyPI
plateau-triage demo --out out/      # runs on a bundled earthquake sample

plateau-triage — Noto demo
  ingested 12 footprints from the bundled Copernicus Noto sample
  coverage: 8 / 12 assessed; 33% unknown
  state breakdown:
    🔴 危険    danger          3
    🟡 要注意   caution         3
    🟢 調査済   inspected       2
    ⬜ 未調査   unsurveyed      4
  outputs:
    triage.geojson  ·  coverage.json  ·  records.duckdb
    support_system.csv  ·  risai_shomei.csv

(危険 / 要注意 / 調査済 / 未調査 are the values actually stored in each record — the Japanese emergency risk-assessment labels for danger / caution / inspected / not-surveyed.)

See the map (the red/yellow/green/grey 3D view above, bilingual JP/EN):

make demo && make viz-data      # build the demo data
make viz-install && make viz-dev   # http://localhost:5173

Field capture (drone / street) and exports for existing systems:

plateau-triage capture --out out/                                   # drone capture → records
plateau-triage export --records out/records.duckdb --out feed.csv --format both

---

How it works

   satellite / public datasets ──┐                  ┌── drone / street capture
        (bulk import)            │                  │   (weak / unseen ⇒ not surveyed)
                                 ▼                  ▼
                        honest per-building records   (missing data ⇒ not surveyed, never "safe")
                                 │ bind to a persistent building ID
                                 ▼
        ┌──────────── record layer (GeoParquet / DuckDB) ────────────┐
        • coverage meter — "how much have we NOT checked?"
        • map feed (red/yellow/green/grey) → the plateau-r3f viewer
        • query — look a building up; list what's still unchecked
        • export — victim-support-system feed + disaster-certificate pre-fill
        • eval — proves, on public data, that it says "unknown" when it should

part	what it does
plateau_triage.record	the one record shape + the four states; the honesty rule lives here
plateau_triage.accrue	bring data in — ingest_batch (datasets) and ingest_capture (drone/street)
plateau_triage.bind	pin each record to a persistent building ID (plateau-id-resolve)
plateau_triage.coverage	the coverage meter + triage ordering (unknown above confirmed-safe)
plateau_triage.store	the record layer on disk (GeoParquet / DuckDB)
plateau_triage.query	look up a building / list what's still unchecked (mounted by caveat-mcp)
plateau_triage.export	feeds for the victim-support system + disaster-certificate pre-fill
plateau_triage.eval	self-proving honesty score (calibration + symmetric penalty)
viz/	the plateau-r3f web viewer (Vite + React Three Fiber, bilingual)

Honest by construction

The core rule — a not-surveyed building never carries a value or gets called safe — is enforced independently at every layer (write, storage, field extraction, export, query), so a regression is a test failure, not a silent slip. The rule, and the seven places it's enforced, are written up in docs/honesty-spec.md. For the data model it is the literal invariant covered=false ⇒ value=null.

For developers

• The record is a frozen, shared contract (contracts/record.schema.json), also projectable onto SOSA + PROV-O and checkable in SHACL — standards alignment, not a new standard. See docs/standards.md.
• pip install -e ".[dev,shacl]" then python -m pytest (118 tests) and cd viz && npm run build.
• Docs: honesty-spec · standards · data-sources · integration.
• Contributing: CONTRIBUTING.md.

License

Code Apache-2.0 (LICENSE); data CC BY 4.0 (LICENSE-DATA, incl. PLATEAU / Copernicus attribution).

This is a screening layer that feeds existing systems — not a replacement for the legal official damage assessment (被害認定調査) a human signs.

---

Yodo Labs · PixelX Inc. / ピクセルエックス株式会社
yodolabs.jp · [e-mail]