Checking a drawing
without pretending to decide.
Our peer-reviewed research reframes automated compliance checking as an auditable decision-support problem: AI assistance combined with explicit evidence and provenance — built on more than 50 assessed South Australian residential cases supplied through DDDI Group and its industry partners.
Automate only what the drawing can actually prove.
Residential compliance review under South Australia’s Planning and Design Code is judged in millimetres, clause by clause, property by property. A model that guesses is worse than useless. So the question we ask of every check is not “can the AI answer it?” — it is “can the drawing prove it, and can a reviewer follow the trace?”
Every reported number traces back to the clause it was compared against and the drawing evidence it came from. The trace is stored as data — a reviewer can follow it without reading code.
Assumptions are data, not hidden defaults. A boundary that doesn’t close returns nothing rather than a fabricated value — failure is a result, not a silent pass.
The system drafts; a person decides. No automated pass or fail is issued at any stage. The output is a review package, not a verdict.
Four stages, one continuous trace.
Specify what to measure, measure it from the geometry, compare it to the controls for this property, then report a package a person reviews. A provenance thread runs under all four — every number can be traced back.
Specify
Planning Code clauses are aligned with the considerations that recur in real assessment outcomes to produce a machine-executable checking specification. What must be measured is fixed here — not generated at runtime by a language model.
Measure
The drawing is read as vector geometry rather than an image, and only the entity types a check requires are recognised. Compliance is judged in millimetres, so the representation has to preserve millimetres.
Compare
The controls governing this specific property are retrieved and presented beside the measurement — with the margin, the clause reference, and the source wording. Zones, overlays and numeric variations differ property by property, so thresholds are retrieved per address before any comparison.
Report
Flagged items return the requirement, the supporting evidence, and the types of adjustment available. No automated pass or fail is issued at any stage — the output is a package a person reviews.
What can be automated, and what stays with a human.
Before a single check runs, every requirement is graded by what the drawing can actually prove. That triage is the difference between a tool that assists and one that overreaches.
Provable from the drawing
Deterministic geometry the drawing already contains. Checked automatically, with the evidence attached.
Provable with one named input
Becomes provable once a single, named human input is supplied. Kept in scope, with the input made explicit — and it becomes Class A once that input is automated.
Judgement a drawing can’t settle
Context or judgement a drawing cannot resolve. Returned as an advisory flag with an explicit threshold — never a verdict.
The rule that applies is a property-level question.
Two addresses on the same street can sit under different zones, overlays and referral triggers. The Compare stage retrieves the controls that govern this address — zoning, overlays, services and market context — and writes each into a register with the reference it came from. Press play to watch one address read.
- Zone provisions
- Overlay controls
- Infrastructure & services
- Market context
- Does not apply
Illustrative pattern — a generated parcel and control register, shown to demonstrate the method, not a live property assessment.
Rules that keep the model honest.
When a language model reads a control, six rules govern what it is allowed to return. Each one closes a specific failure mode.
Use only retrieved text
Blocks rules from another jurisdiction leaking in.
Do not infer or invent
Blocks fabricated thresholds.
Say “not specified”
Leaves a reviewable gap instead of a silent omission.
Keep source wording verbatim
Preserves the exact control a reviewer will check against.
Return each control as its own record
Stops multiple controls being averaged into one.
Emit structured output only
Keeps every result machine-checkable and traceable.
The review package.
“Auditable” is not a claim — it is a payload. Six things accompany every flagged item, so a reviewer can check the result without trusting the tool.
“A number you cannot trace is a number you cannot use.”
AI-Assisted Decision Support for Drawing-Based Residential Compliance Review
Integrating Large Language Models and Computer Vision. Peer-reviewed and amended over four revision rounds, and presented at WSBE26 — the World Sustainable Built Environment Conference 2026, Melbourne Convention and Exhibition Centre.
More than 50 assessed South Australian residential cases, each pairing the plan drawing with its official assessment outcome — supplied through DDDI Group and its industry partners.
Yupeng Zhang (presenter), Ruidong Chang, Xingyue Fang, and William Jiang.
Setting: the South Australian Planning and Design Code.
Compliance you can audit.
Talk to us about applying drawing-based compliance review to your assessment pipeline — or explore the wider Future Housing Intelligence Program it belongs to.