From public geometry to an inspectable comparison

Road Risk asks whether a transparent combination of road geometry, physical relationships and explicit assumptions can help researchers compare road sections without pretending to know real-world safety outcomes.

Read the method Open the App

Question and hypothesis

What is being tested?

Research question

Can road geometry, physical demand and controlled scenario assumptions be combined into a reproducible comparative output that remains transparent about missing evidence?

Hypothesis

Tighter geometry, higher speed demand, lower assumed friction and longer stopping demand should change the model output in predictable directions when other inputs are held constant.

The hypothesis concerns model behaviour. It does not establish collision probability, a safe threshold, an official road classification or a driving recommendation.

Variables and controls

Separate what is observed, calculated and selected

Observed inputs: Public road centreline geometry and available map tags.
Derived values: Length, heading change, curvature, radius, lateral demand and stopping-distance comparison.
Selected controls: Weather, surface, lighting, behaviour, vehicle and speed assumptions.
Held constant: Road, model version and all settings except the variable being examined.

Implemented system

One workspace connects map, calculation and evidence

Road selection

Select public geometry and preserve source details.

Comparative calculation

Run the chosen model under explicit settings.

Local evidence

Inspect 100 m sections, routes, areas, Maths and geometry diagnostics.

Review and replay

Export source-labelled records and reopen supported artifacts in the App.

What exists now

Software evidence is stronger than empirical evidence

Current model status

Model 2.0.0 is the public default

Compare model versions

Model 2.0

Public default

Empirical decision: not assessed

Model 3.2

Controlled preview

Empirical decision: insufficient evidence

Model 3.3 geometry candidates

Diagnostic only

Empirical decision: insufficient evidence

Software verification and deployment status are reported separately from empirical evidence. No model output is an official safety rating or collision prediction.

The repository contains deterministic fixtures, model goldens, trace checks, geometry tests, import/export contracts and historical reviewed model-output records. The frozen Irish empirical corpus, final temporal holdout and independent reviews are not complete.

Limitations and next work

What the project still needs

Exposure data

Measured traffic exposure and stable road matching are required before outcome-aware evaluation.

Collision evaluation

Spatially grouped development analysis and an untouched future holdout remain required.

Independent review

Geometry, statistical and transport-domain reviews are not yet recorded.

Operational verification

Real-device, assistive-technology and live-production checks remain separate release gates.