# Epistemic Risk Classes
*When a geospatial agent must escalate from procedure → methodology*
## 1. Purpose
This document defines **where** geospatial operations carry *epistemic risk* — meaning situations where the correctness of the result depends on the **methodological justification**, not merely the computational outcome.
These risk classes are the conditions under which the agent must **elevate reasoning from “do” to “justify.”** They activate the epistemic escalation described in `AGENT_EPISTEMOLOGY.md`.
---
## 2. The Four Primary Classes of Epistemic Risk
| Class | Core Risk | Why It Matters |
|-----------------------------------|------------------------------------------------------|--------------------------------------------------------------------------|
| 1. CRS / Datum Risk | Misinterpretation of space or elevation | Can distort scale, area, or topology without visible failure |
| 2. Resolution / Resampling Risk | Loss or mutation of signal | Results remain “numerically valid” but become scientifically meaningless |
| 3. Topology / Hydrology Risk | Structural discontinuity in surfaces or networks | Breaks flow direction, drainage, adjacency, routing |
| 4. Aggregation / Statistical Risk | Unjustified mathematical combination of spatial data | Converts true spatial signals into misleading summaries |
These four classes cover *all* categories where GIS can silently produce wrong science.
---
## 3. Subclasses Within Each Category
### 1. CRS / DATUM RISK
* Horizontal CRS mismatch (projection vs ellipsoid assumption)
* Vertical datum mismatch (geoid vs ellipsoid height)
* Mixed geodetic epochs (tectonic movement)
* Loss of linear vs areal property preservation
### 2. RESOLUTION / RESAMPLING RISK
* Pixel aggregation (nearest vs bilinear vs cubic)
* Over-smoothing or staircasing
* Mixed resolution mosaics
* Derived terrain artifacts from interpolation
### 3. TOPOLOGY / HYDROLOGY RISK
* DEM hydrological conditioning
* Sink/breach decisions
* Snap-to-stream tolerance
* Breaklines vs implicit surfaces
### 4. AGGREGATION / STATISTICAL RISK
* Averaging across nonlinear terrain proxies
* Zonal statistics on mixed-resolution rasters
* Spatial joins that assume stationarity
* Ignoring variance or uncertainty in combined layers
---
## 4. Silent Failure Modes (representative)
| Class | Silent Failure Example |
|-------------|----------------------------------------------------------------------------------------------------|
| CRS | A DEM “looks fine” but distances shrink by 8% due to wrong projection; watershed boundaries shift. |
| Resolution | Aggregation smooths slope rasters, reducing hazard classification severity without any error flag. |
| Hydrology | Flow routing selects a false pour point because microtopography was lost during conditioning. |
| Aggregation | Zonal mean temperature hides extreme microclimate pockets → false suitability conclusions. |
None of these *crash*.
They “succeed” — but produce invalid science.
---
## 5. Epistemic Escalation Triggers (internally framed)
Escalate to methodology whenever you (the agent):
### CRS / Datum Risk
* are unsure whether the target CRS preserves the scientific property required (distance, area, hydrology, azimuth integrity)
* cannot confirm vertical datum compatibility
* detect a likely mismatch between input and target assumptions
### Resolution / Resampling Risk
* realize that interpolation or kernel choice could change interpretation of terrain or gradients
* are unsure whether native resolution held ecological or hydrological meaning
### Topology / Hydrology Risk
* suspect the DEM’s drainage structure may not be topologically sound
* are taking actions that implicitly assume “surface connectivity”
### Aggregation / Statistical Risk
* combine values whose meaning depends on spatial variance or sampling theory
* are unsure whether aggregation obscures signal vs noise
---
## 6. Boundary Conditions (non-escalation)
Escalation is *not required* when:
* Data is already in target CRS and verified compatible.
* Resampling is identity-resolution (copy) or for pure display.
* Hydrology/topology is not inferred from the data.
* Aggregation is not interpretive (e.g. metadata concatenation, structural staging ops).
This reinforces that escalation is **contextual**, not universal.
---
## 7. Relationship to Methodology
After a risk class is triggered, the agent should *not immediately call a GDAL tool.*
Instead, it should seek a **methodological scaffold** appropriate to the risk:
| Risk Class | Methodology Type |
|-------------|-----------------------------------------|
| CRS / Datum | Geodetic justification |
| Resolution | Sampling / interpolation justification |
| Hydrology | Topological continuity justification |
| Aggregation | Statistical defensibility justification |
Methodology comes **before** execution.
---
## 8. Forward-Compatibility
These classes are intentionally broad because future models may:
* invent new hydrological correction heuristics,
* propose projection strategies not in today’s textbooks,
* discover new sampling-aware transformations,
…yet these four epistemic gateways will *still* define
**when** justification is required.
This preserves scientific rigor *while remaining compatible with future discovery.*
