Case notes
Method, number, limit
Each piece of flagship work below says what we did, the headline number, its limit, and what we will not claim. Follow the link for the full public case note.
1. Statewide signal evidence
Method. Configuration sheets go to AI vision, which reads the detector maps. These are matched to roads within 160 m and pass one geographic and value QA gate.
Number. 4,527 of 4,588 volumed signalised intersections have per-approach evidence (98.7%).
Limit. This is coverage of volumed signalised intersections, not of every signal in Victoria. One vision model returned systematic false negatives; re-reading recovered 56 of 64 cases, and each model's miss pattern is now recorded.
2. Directional volume for every road
Method. Measured counts are used where they exist, signal-intersection evidence fills the urban network, and a model covers the rest. The total is checked against an independent vehicle-km envelope.
Number. 1,267,052 road sections and about 71 billion vehicle-km.
Limit. Official volume sources disagree by more than many accuracy targets assume. The disagreement table is published, so any accuracy claim is scoped honestly.
3. A worked before/after example
Method. A comparison-group before/after study. An earlier engine refused to report a low-count study at all, so the lab rebuilt it to publish the estimate and let the interval carry the weakness.
Number. Crash modification factor 0.75, 95% CI 0.30–1.85.
Limit. The interval includes 1, so "no effect" is not ruled out. The result is consistent with a reduction, but it is not evidence of one.
4. Speed zones and aggregation bias
Method. Six NSW sites, weighted difference-in-differences against control roads, and a rolling coefficient of variation. Connected-vehicle speeds are checked against two instrumented research cars.
Number. Six speed-zone changes evaluated.
Limit. The note presents the method. Site-level effect sizes are not presented as findings.
5. Skewed intersections
Method. Every Victorian intersection is classified by approach angle into three bands and joined to crash history.
Number. 34,220 skewed intersections.
Open. The data are public and the code is open source.
All six areas of expertise
- System integrity
- Statewide network evidence
- Spatial road risk
- Speed, evaluation and connected vehicles
- Crash data systems and analysis
- Verified AI and machine learning
Want to know where this work has been used? See Work relied on.