Board briefing · September 2026
A plain-language walk through the Australian Property Investment, Accommodation and Operations Ontology (version 0.1.0): what it is, the shape it takes, how the pieces connect, and what it lets us do.
Status: release candidate · all 15 test questions pass · every term traced to a source · prepared from the generated Semantic Atlas and validation reports
Bottom line
The ontology names every kind of thing the business deals with (a room, a head lease, a council registration, a yield claim), states how those things relate, and records where each fact came from. It is the common language beneath the property-intelligence platform and any AI assistant we ground on our research.
Three things the board should know
Why we built it
Rooming-house gross yield is quoted anywhere from 5% to 15%. Coliving break-even is quoted as both 6–12 and 12–18 months, by the same publisher. Room-by-room letting is said to both lower and raise vacancy.
Four residents in Queensland trigger registration and accreditation; four persons in Victoria trigger an operator licence and council registration; under nine bedrooms in Victoria can avoid a planning permit. The word “rooming house” itself means something different in each place.
An operator is the owner’s tenant and the resident’s landlord at the same time. An owner-occupier can also be a provider. Spreadsheets and documents cannot hold that cleanly; a map of roles can.
The ontology turns 31 web pages, 14 summaries and two research reports into one structure where every statement carries its source, its authority level and its jurisdiction.
The shape
Where the connections concentrate
What it means in practice
Most of the other 90 concepts are deliberately simple and connect through one of these three.
How it is put together
100% of terms have a definition and at least one evidence record or written rationale. Zero references to unknown sources.
Classification values that sources argue about (strategies, models, risk categories) live in controlled lists, not in the structural hierarchy. The hierarchy is reserved for things that are always true, such as “a sublease is an agreement”.
The model files, the 197-diagram atlas and this report are generated from the same bundle. Edit the build script, not the outputs.
The controlled vocabulary
Why lists rather than definitions
When a source says “coliving” or “HMO”, we record that value. We do not attempt to define the boundary between them, because the sources do not agree and the business combines them freely (a rooming house, run under a master lease, held in a company).
Examples
How the pieces connect: a worked example
This is the pattern that the whole model repeats: a thing, the parties around it in named roles, the agreements that bind them, and the rules that apply. Every arrow can be asked about.
Keeping disagreement honest
The model stores each figure with its publisher, date, jurisdiction, basis and authority level. It never averages them or declares one correct. That is a deliberate decision (DEC-008), and it is what makes the AI’s answers auditable.
The rules that switch on and off
These are the single most decision-critical facts in the corpus, so each is a first-class record with its kind, comparison, value, unit and effect (DEC-007). The compliance and approvals questions read them directly. * Victorian Clause 52.23 limits come from a secondary research report and are flagged as not yet verified against the planning scheme (ISSUE-005).
What it lets us do
| Property (example record) | Operating model | Authorisation required | Status |
|---|---|---|---|
| Brisbane six-room purpose-built rooming house | Owner-operated | Accommodation service accreditation (QLD) | missing |
| Brisbane six-room purpose-built rooming house | Owner-operated | Residential service registration (QLD) | pending |
| Melbourne five-bedroom house under master lease | Master lease / rent-to-rent | Building approval / classification | missing |
| Melbourne five-bedroom house under master lease | Master lease / rent-to-rent | Fire-safety certification | missing |
| Melbourne five-bedroom house under master lease | Master lease / rent-to-rent | Council premises registration (VIC) | expired |
Output of test question CQ-015, “which properties have a lawful-use authorisation that is missing, pending or expired?”, run against the worked-example records. Two example properties, five gaps.
How the answer is assembled
The same machinery answers all 15 questions. Nothing here is hand-written; it is the structure doing the work.
Rules have dates
How strong is the evidence?
Promotional yield claims are catalogued as claims, never accepted as fact. Legal and tax detail that appears only in a synthesis report is recorded but flagged.
Does it work?
| ID | Question, in plain terms | Rows | Result |
|---|---|---|---|
| CQ-001 | Which approvals does an operating model need in a state, and who regulates them? | 25 | pass |
| CQ-002 | What occupancy or size threshold triggers or exempts a rule? | 14 | pass |
| CQ-003 | Under an agreement, who plays which role and who carries the rent risk? | 30 | pass |
| CQ-004 | What assumptions went into a financial scenario and what came out? | 13 | pass |
| CQ-005 | What are the characteristics of each ownership structure? | 9 | pass |
| CQ-006 | Which accommodation type, operating model and tenants does each strategy rely on? | 10 | pass |
| CQ-007 | Which risks apply, and what mitigation is recorded? | 8 | pass |
| CQ-008 supporting | Which sections must a business plan or feasibility study contain? | 30 | pass |
| CQ-009 | Which data source supplies which property attribute, at what cost and cadence? | 18 | pass |
| CQ-010 supporting | Which lease types put which costs on landlord versus tenant? | 9 | pass |
| CQ-011 | Who claims what figure for a metric, and on what authority? | 8 | pass |
| CQ-012 supporting | Which tax rules apply to which property class from which date? | 9 | pass |
| CQ-013 supporting | Which services, fees, workflows and KPIs run a management business? | 7 | pass |
| CQ-014 supporting | Which agency business models exist and how are capabilities sourced? | 5 | pass |
| CQ-015 supporting | Which properties have a missing, pending or expired authorisation? | 5 | pass |
Nine core questions and six supporting ones (marked). “Rows” is the number of answer rows each returned; every count matched the expected value set after inspecting the answers.
Not done: no automated logic reasoner has been run (policy does not require one); no independent expert review; synthesis-only legal and tax claims not verified against legislation.
What is unresolved
Yield, vacancy and break-even figures disagree. Every claim is kept with its source; no single “truth” is asserted.
“Rooming house” means different things in each state. Modelled as one broad concept with state-specific triggers.
Key regulatory detail and the 2027 tax changes come from a secondary report. Flagged; primary verification still required.
Spanish, Canadian and European concepts kept as generic patterns only; no Australian applicability asserted.
A manifest inconsistency and a modelling shortcut, both recorded and handled.
Seven web pages failed to download. Victorian coliving economics and multi-lease lending are under-represented.
Each item is recorded in the model’s own issue register with the evidence for and against, an authority analysis and how it was handled. “Handled” here means the disagreement is represented faithfully, not that a winner was chosen.
What it would take to release
Endorse version 0.1.0 as the working vocabulary for AI grounding and platform design now. Treat any legal or financial figure it returns as “sourced claim, verify before acting” until the four items on the left are closed, at which point it can be promoted to a released version.
Released versions are immutable; corrections come as a new version number, so downstream systems can rely on stable meaning.
Appendix
Source material: documentation/semantic-atlas.md (generated, 197 diagrams), validation/validation.md, validation/cq-results.json, sources.yaml, decisions.yaml, model.yaml. All counts in this report are read from those files.