A Public Data Atlas
of the Federation

01Dr Non Arkaraprasertkul

Most important thing first: this is a gift.

Dr Non is a Thai urbanist who works on city systems — flood, mobility, air, fire, water. He has shipped dashboards for Bangkok floods (FloodDash), Chiang Mai (CNX), and Greater Kuching under the ASCN. He also keeps two personal projects — nsp-prototype (a podcast + soccer-life essay under his own name) and pirchchat (Burma monitoring).

The Malaysia project extends three boards he has already shipped. Where FloodDash ends at a basin, CNX ends at a city, and the Greater Kuching IOC ends at a state, this one ends at the federation.

"A good dashboard tells you what the room is doing right now. A great dashboard tells you what the room cannot tell you. This one does both, in three languages."

— Dr Non Arkaraprasertkul, project notes, September 2026

a = sin²(Δφ / 2)
  + cos(φ₁) · cos(φ₂) · sin²(Δλ / 2)
c = 2 · atan2(√a, √(1−a))
d = R · c                R = 6,371 km

for each facility f ∈
  {hospitals, clinics, schools}:
  km[c, f] = haversine(c, f)
  nearest[c] = argmin_f km[c, f]
  if km[c, nearest_hospital] > 20:
    people_>20km += n
total_>20km = Σ people_>20km

3.3 Freshness grading

Observation feeds carry the reading or reference-period time. Event lists carry the upstream source-check time. These meanings are labelled separately. Age against cadence decides status:

grade(f) =
  loading   if !f.fetched
  error     if f.error
  nokey     if error matches /no .*key/
  stale     if f.obs > now + 5m
  stale     if age > 4 · cadence
  delayed   if age > cadence
  live      otherwise

A 200 OK with yesterday's numbers is not "live".

3.4 Trilingual interpolation

Bahasa forecast strings are matched against a small phrase table in feeds.js and replaced with EN or 中文 equivalents. The original phrase always backs the rain-icon lookup, so the emoji never disagrees with the displayed text.

3.5 Air quality

PM2.5 for the board’s 21 cities comes from Open-Meteo CAMS model estimates in µg/m³. The concentration display uses legacy bands at 12 / 35 / 55 / 150 µg/m³; it is not a current EPA AQI calculation. DOE MyEQMS separately supplies observed Malaysian API/IPU station indices, checked with the shared three-hour currency rule. These signals have different scales and cannot be subtracted.

3.6 Reach map (visual)

FACEBOOK_PAGE_ID            = 123456789
FACEBOOK_PAGE_ACCESS_TOKEN  = EAAxxxxxx...

Example wire shape

GET /api/facebook

200 OK
{
  "source": "Facebook Graph API",
  "configured": true,
  "posts": [
    {
      "id": "123_456",
      "message": "Today at 14:32 MYT...",
      "time": "2026-09-25T06:32:00+0000",
      "url": "https://facebook.com/...",
      "image": "https://scontent...",
      "reactions": 432,
      "comments": 41,
      "shares": 18
    }
  ]
}

Failure mode

If authorized access is allowed but the token has expired or the page is unpublished, the relay returns 424 (Failed Dependency) with error set. The press tab uses Google News independently.

What the dashboard does not claim

• Hospital locations are reference data. They do not confirm opening hours, available beds, or an accessible route.
• Reach distances are straight line. Real journeys by road or river are longer.
• Census 2020 and HIES 2019 are historical.
• CAMS air quality is modelled. DOE API/IPU is observed at stations; neither establishes conditions at every address.
• Ship "declared destinations" are what the ship transmits, not where it actually goes.
• Rain radar is 10-minute frames.
• Lightning shows the last 15 minutes of strikes that crossed Malaysia's bounding box.
• The optional Facebook adapter requests public Page posts and aggregate reaction/comment counts, not DMs or comment bodies. It is not a current dashboard source.

What it does claim

• If a river gauge reads DANGER within the last 45 minutes, that is the operator-reported category for that station.
• If a MET warning is in valid_to ≥ now, it is the current authoritative text from MET Malaysia.
• If the verdict says Not enough data, coverage is incomplete or a reading cannot be used; upstream failure and dashboard defects are both possible.

score = category      DANGER 100 · WARNING 60 · ALERT 30
      + rising trend              +25
      + climbing ≥ 0.3 m/h        +15
      + rain at the station       +10
      + rain in state forecast    +10
      + MET warning in state      +15
      + low-ground people nearby  +min(20, pop / 2000)

Worked example: the morning of 27 September 2026

At 11:00 MYT the relay held 1,306 stations with 503 current readings (newest 12 minutes old). Five MET warnings were in force — strong winds and rough seas naming Kelantan, Pahang, Perak, Terengganu, Sabah and Sarawak, plus thunderstorms over Kelantan, Sabah, Sarawak and Labuan. The queue ranked:

#	Gauge	Evidence
1	Sg. Bidor di Changkat Jong, Hilir Perak — DANGER 3.92 m against a 3.5 m line	rising · 1 mm at station · MET warning in Perak · 38,278 on low ground nearby
2	Sg. Bernam di Jambatan SKC, Hulu Selangor — DANGER 19.21 m against 19.1 m	receding · reading 42 min old · verify before standing down
3	Sg. Luit di Kg. Subuh, Maran, Pahang — WARNING, steady	MET warning in Pahang
4	Sg. Galas di Limau Kasturi, Gua Musang, Kelantan — ALERT, receding	25 mm at station · MET warning in Kelantan

Hilir Perak holds 141,959 people (Census 2020), 38,278 of them on low ground by rivers — and 0% live over 20 km from a hospital. So the morning's answer was specific: the danger in Perak was exposure, not distance to care. A different district with the same gauge would have ranked an evacuation-hard zone alongside it. That distinction — what kind of danger, not just where — is the whole point of the joins.

"Raw data tells you what the room is doing. Joins tell you what to do about it. If an insight cannot change someone's next hour, it does not belong on this board."

— Dr Non Arkaraprasertkul, insight-layer notes, September 2026

Method and authorship

Project notes, September 2026. The worked example is historical and its raw response archive is not supplied. Formulations are implemented in public/js/insights.js, rendered trilingually in the briefing tab, and held to account by tests/insights.test.mjs — nine tests covering ranking, evidence convergence, climb memory, zone logic, Bahasa forecast parsing, and trilingual copy coverage.

for each HRSL cell c:
  for each facility f in [hospitals, clinics, schools]:
    km[c, f] = haversine(c, f)
  nearest[c] = argmin_f km[c, f]
  if km[c, nearest_hospital] > 20:
    people_>20km += population(c)
total_>20km = Σ people_>20km

Reach analysis metrics

• People within 20 km: Σ population(c) for all cells where the nearest hospital ≤20 km
• People beyond 20 km: Σ population(c) for all cells where the nearest hospital >20 km
• Percent beyond 20 km: total_beyond / total_population × 100
• Average distance to nearest hospital: Σ population(c) × km[c, nearest[c]] / Σ population(c) (population-weighted)
• District reach score: no composite score is calculated; distance and population exposure remain separate

Applications in evacuation planning

The briefing flags districts with a hospital-distance gap when a current hazard names their state. This is a review priority, not an evacuation order or route assessment. Straight-line distance does not establish whether a hospital can be reached during a flood.

Data sources validation

• HRSL grid cells: no road-density validation study is supplied; the former R² claim is unsupported.
• Facility locations: Overture reference records; the tests check plausible counts and the Semporna location correction, not a complete official-register audit.
• Population totals: tests require the national baked-grid total to fall within 3% of Census 2020. That reconciliation does not validate each cell or census tract.

Limitations

• Static data ignores seasonal migration (students, workers)
• Road network distances not calculated (approximate local kilometre projection)
• Facility accessibility (road condition, traffic, language) not considered
• Weather-dependent routes not factored