Statewide signal evidence from 4,870 sheets

Scale

The numbers

~4,870configuration sheets (spreadsheet and scanned PDF)

4,588signalised intersections with volume data

4,527with per-approach evidence (98.7%)

61residual sites, each with a stated reason

Method

How it works

Configuration sheets→AI vision reads detector maps→Match to roads within 160 m→One geographic + value QA gate→Per-approach evidence

Scanned sheets run 10–20 pages with no text layer, and the detector map is usually on page 2 or 3, not the cover. Where no configuration exists, a site-name geometry fallback matches cross-street names to mapped roads and passes the same QA gate.

Evidence mix

Where the evidence comes from

Scanned detector maps~2,070

Signal-group evidence~1,410

Spreadsheet sheets~1,510

Per-road detector data357

Site-name fallback152

Sources overlap; the union is 4,527 sites.

Integrity

What went wrong, and what we changed

A model that said "no map" when there was one. One vision model systematically returned false negatives. Re-reading with a reference model recovered 56 of 64 cases. We now record each model's miss pattern and do not trust a single model's "no map".

A review screen that showed the wrong page. Thumbnails showed the cover sheet, so good maps looked unreadable. Review now renders pages 1–6.

Limits

What we will not claim

98.7% is coverage of the volumed signalised intersections, not of every signal in Victoria. The 61 residual sites are listed by reason: oversized files, sheets missing from the published download, and sites with no mapped road within 160 m. Per-site volumes are not published on this site.