SinkAlert — Overnight Research Files

Synthesis Final Reports

🧊

phase-b-rf-status.html · Phase B RF v1.56 FREEZE

Canonical · Jul 29, 2026 · OOF AP ~0.945 · F1 ~0.957 · research only

FROZEN web default (12 labels). Catalog growth 54 BMR / 49 unique (Round 11) is not promoted — dual-gate held. Spec: docs/100-phase-b-rf-v156-FREEZE.md.

Open page Metadata MHEWS next

🛡️

oic-insurtech-2026.html · OIC InsurTech Award 2026

Pitch · Jul 29, 2026 · Challenge #3 Disaster risk · BMC PDF · research RF honest

OIC + CIT award prep. Category #3 disaster-related risks for insurers: underwriting awareness, portfolio screening, claims triage, gov coordination. Maps 5 judging perspectives. RF phase_b_rf_v1_56 research-only. Doc: docs/112-oic-insurtech-award-2026.md.

Open pitch BMC print RF freeze

🛰️

phase-b-gistda-brief.html · Phase B GISTDA Brief

Partner brief · Jul 29, 2026 · white/blue · Thai + EN · printable · research only

GISTDA-facing brief. Catalog n_pos=54 / 49 unique vs freeze 12-train v1_56; why denser-label + spatial CV dual-gate held; CopDEM/BMA/PSI/GW/OSM/GOB stack; Contabo demos; MEA/MWA/LiCSBAS/GISTDA collaboration ask. operational_use=false.

Open brief Freeze status GISTDA ↗

🗺️

phase-b-research-map.html · Research compare map

Spike · Jul 29, 2026 · MapLibre swipe / dual-pane · research only

GeoLibre ideas 1–2 (lite). Static HTML compare: frozen RF (v1_56 · 12 labels) vs news/sample incidents; catalog growth 54 BMR / 49 unique noted on banner. Not COG, not operational CAP.

Open map Freeze status News GeoJSON

🔴

SINKALERT_OVERNIGHT_CRITIQUE.md

17 KB · 407 lines · Jun 29, 2026

CRITICAL READ. Ginnie's brutal verdict (5.5/10 — NOT READY). Cross-examination of every claim, contradictions with Babigon, 10 judge questions, Week 1 fix plan (+17 points).

⬇ Download GitHub

🟡

SINKALERT_OVERNIGHT_SYNTHESIS.md

16 KB · 291 lines · Jun 29, 2026

Babigon's unified synthesis: Top 10 data sources, Top 10 papers/tools, 4-phase integration roadmap (260 hrs), pitch deck recommendations, SWOT analysis.

⬇ Download GitHub

BDI Requirements gap analysis

Contabo nginx returns 403 for several legacy HTML names (bdi-requirements, strategy-*, yolo-research, research-literature, research-thaillm, sinkhole-causation). Mega-menu links point here (or to live pages) instead. Source HTML remains in the repo for local use.

📋

docs/004-bdi-requirements-gap-analysis.md

Canonical gap analysis · GitHub

BDI / DEPA data requirements vs what SinkAlert can show honestly today. Prefer this doc over the Contabo-blocked HTML mirror.

Open on GitHub Impact Analysis Master Research

Babigon Global Data Hunt — 4 Rounds

🔵

babigon_r1_data.md

18 KB · 471 lines

Round 1: 91 potential data sources discovered — open + paid + academic. Cast the widest net across Zenodo, NASA, ESA, USGS, BGS, commercial satellites, insurance databases, Thai government portals.

⬇ Download

🔵

babigon_r2_ranked.md

18 KB · 444 lines

Round 2: Top 20 verified — URLs tested, data schemas inspected. Ranked by composite score (accessibility + predictive value + cost). All top 10 are FREE.

⬇ Download

🔵

babigon_r3_deep.md

15 KB · 391 lines

Round 3: Deep dive on top 5 sources. Actual data fields, feature mapping to SinkAlert, F1 gain estimates, sample queries, integration pseudocode.

⬇ Download

🔵

babigon_r4_integration.md

13 KB · 342 lines

Round 4: Step-by-step integration order. Cost breakdown (everything free for research). Time estimates. Expected F1 improvement per source (+0.02-0.08 each).

⬇ Download

Ginnie Academic & Code Deep Research — 4 Rounds

🟢

ginnie_r1_papers.md

13 KB · 268 lines

Round 1: 19 key papers with DOIs. Sinkhole prediction, InSAR+ML, pavement CV, groundwater-subsidence, karst mapping, physics-informed NN, graph neural networks, foundation models.

⬇ Download

🟢

ginnie_r2_code.md

13 KB · 300 lines

Round 2: 24 production-ready tools. GitHub repos (stars, activity, license), HuggingFace models, QGIS plugins, SAR processors. Includes qgis-mcp (161★) evaluation.

⬇ Download

🟢

ginnie_r3_benchmark.md

10 KB · 244 lines

Round 3: Benchmark against 3 best published models. Where SinkAlert wins (multi-modal fusion) and falls short (no early warning, no temporal validation). 3 weakest links identified.

⬇ Download

🟢

ginnie_r4_roadmap.md

14 KB · 328 lines

Round 4: Top 10 actionable improvements ranked by impact. 3 novel ideas no other team will have. AWS services to add. Per-dimension pitch improvements with score estimates.

⬇ Download

Training YOLO Baseline Results

🤖

yolo_baseline_results.md

501 B · 14 lines

CPU training baseline: mAP@50 = 0.25 (failed). Needs SageMaker GPU on full 7,598-image dataset.

⬇ Download

🌙 SinkAlert Overnight Research