Model & Method
Nothing here is a black box. Each model states what it is, what it learned from, how well it performs, and where it should not be trusted.
Standing inferences
8
Mean confidence
0.74
Domains fused
5
Hydrology, geoscience, bioscience, remote sensing, social
Models in use
| Model | Approach | Trained on | Performance | Weakest where |
|---|---|---|---|---|
| Proxy regression (pH + EC → dissolved copper) | Gradient-boosted regression | 98 paired laboratory and in-situ samples, 2022–2026 | R² 0.83, RMSE 0.19 mg/L | Above 1.2 mg/L copper and during first-flush wet-season events |
| Source separation (chemistry + eDNA) | Unsupervised clustering with chemistry constraints | 9 eDNA libraries, 98 water samples | Silhouette 0.61; SN5 separates on every run | Single dry-season eDNA epoch; no wet-season replicate |
| Recovery forecast | State-space trend model with treatment term | 9 quarterly rounds at SN5, one intervention record | Backtest MAPE 11% at 2 quarters, 26% at 6 quarters | Treatment media exhaustion timing is unobserved |
| Revegetation prioritisation | Multi-criteria ranking with learned weights | 29 soil samples, 6 plant records, 2 survey epochs | Rank agreement with expert panel: 5 of 6 | Substrate depth inferred from surface change, not measured |
Known data gaps
- No wet-season eDNA epoch — source separation is dry-season evidence only.
- Remote sensing is the oldest domain at 5 months; change rates are two-epoch steps, not trends.
- Dry-season irrigation abstraction is modelled, not metered.
- Zone A crop tissue sampling is below the threshold needed to issue guidance.
- Zones D and E are suppressed at n<10, so no zone-level social signal exists for them by design.
Human-in-the-loop boundary
Machine
Detects change, correlates across domains, forecasts trajectories, ranks by materiality, drafts plain language, and schedules the conversation.
Human
Decides. Weighs community priority against cost and risk, approves what is published, and owns the consequence. No decision in this platform executes without a person recording why.
Remote sensing epochs
- 2025-09-14Drone LiDAR DEM + 5-band multispectral
- 2026-02-27Drone LiDAR DEM + 5-band multispectral
eDNA libraries
- SN5 — AMD-impactedAcidithiobacillus spp.
- SN5 — AMD-impactedLeptospirillum spp.
- SN6 — TransitionalFerrovum spp.
- SN4 — TransitionalGallionella spp.
- SN3 — TransitionalSideroxydans spp.
- SN1 — Clean waterFlavobacterium spp.
- SN2 — Clean waterPolaromonas spp.
- SN7 — Clean waterRhodoferax spp.
- SN8 — Clean waterLimnohabitans spp.
Geoscience site classes
- ControlPLI 0.4
- Tailings Storage FacilityPLI 6.8
- Waste Dump APLI 4.9
- Waste Dump BPLI 5.6
- Bugtong FarmPLI 1.9