Phase B Random Forest v1.56

Official freeze of the open-data RF susceptibility model. Use this page as the human reference; agents start from docs/100-phase-b-rf-v156-FREEZE.md.

Live governance verify

Checking Contabo meta…

Catalog counts bind from data/news-sinkholes-2024-2026.json (not hardcoded marketing). What we do NOT claim · Partner GIS ask · Duty brief

Headline metrics (OOF)

—

Average precision (OOF)

—

Best F1 (OOF)

—

LOPO AUC (mean)

operational_use = false · cap_auto_issue = false · Screening only (no routing authority) · CAP drafts ≠ CAP issuance · Never cite in-sample train scores · Never auto-issue CAP from this layer.

Model: — · OOF AP — · F1 — · research only · not CAP

Model card

Field	Value
Model ID	`phase_b_rf_v1_56`
Algorithm	sklearn RandomForest (500 trees, depth 4) — unchanged
Labels	12 BMR news positives · 60 negatives (frozen web model `v1_56`)
Catalog (bound from JSON)	— BMR events · 49 unique points (<100 m co-located flagged) — catalog growth, not the frozen web weights; `v1_89`–`v1_95` dual-gate held; spatial-block CV in `docs/117`
Validation	LOPO (freeze 12) · denser LOPO unique (49) · spatial k-means / grid (`v1_96`, diagnostic)
Best F1 @ ~0.30	11 TP / 1 FN / 0 FP
Hard FN	Sai Noi MEA cable-pit (`nbi-sainoi-2025-07-19`, p≈0.08)
Map layer	Dashboard → “RF susceptibility (research)”
Artifacts	`data/phase-b-rf-metadata.json` · GeoJSON grid · `models/phase_b/rf_v1_56.joblib`

Enterprise pilot use (honest)

Allowed: enterprise pilot screening (research-only) to help prioritize patrol / inspection attention when verified news and citizen reports are thin; human decisions remain required.
Not allowed: routing authority, automatic CAP issuance, citizen-facing “risk now” scores, or claims that the layer is deployable susceptibility for BMA dispatch. CAP drafts ≠ CAP issuance.
Toggle on Dashboard: “RF enterprise pilot screening (optional)” · operational_use=false · see also GISTDA brief.

Why frozen

Open-feature trials through v1.88 (Geofabrik, GHSL, GSW, HydroRIVERS, WorldPop, OSM power, peri-urban) failed dual promotion gates.
Denser-label retrains v1_89–v1_95 (catalog → 54) and unique-point / spatial-block v1_96 failed dual gates — holding v1_56 is scientifically honest (see docs/117).
Sai Noi needs MEA underground GIS (or denser utility labels) — OSM towers ≠ manholes.
Further open churn without partner data (MEA/MWA/LiCSBAS/GISTDA) is not productive.

What unlocks v1.57+

MEA Map API / underground cable–manhole GIS
MWA pipe age / deterioration GIS
Citywide LiCSBAS / numeric PSI GeoTIFF (not JPEG maps)
Spatial-block CV now run on 49 unique points (v1_96): k-means AP ~0.38 / AUC ~0.75 — held vs v1.56; next lift needs MEA/MWA/LiCSBAS (see docs/117)

GISTDA partner brief

White/blue partner brief for GISTDA GIS / remote-sensing reviewers — catalog 54 events / 49 unique vs freeze 12, why denser LOPO + spatial CV held dual gates, Contabo demo links, MEA/MWA/LiCSBAS/GISTDA collaboration ask. Printable CSS · research only.

Open GISTDA brief →

Research compare map

Static MapLibre swipe / dual-pane: frozen RF susceptibility vs news / sample incidents. Research only — same freeze rules as this page.

Open research compare map → GISTDA partner brief →

Next product step

Science freeze complete. Ops focus: Contabo MHEWS — PM2 sinkalert-mhews on port 8780 + nginx /api/mhews/, then SSO before government rollout. See MHEWS overview and docs/101-next-mhews-contabo.md.

Spec: docs/100-phase-b-rf-v156-FREEZE.md · Agent main ref: docs/000-AGENT-MAIN-REFERENCE.md · Frozen 2026-07-29 · SinkAlert research