From published law to a cited, checkable knowledge base.

The pipeline

1. 1

   Fetch

   Download the official file from the primary publisher (legislation.gov.au for the Act, Regulations and instruments) and record its provenance: URL, retrieval date, SHA-256, and the exact compilation/version.

2. 2

   Extract

   Pull the text and the Schedule tables out of the PDF (the website's text view omits the tables). The raw, layout-preserving text is kept as the machine-diffable substrate.

3. 3

   Structure

   Map the extracted rows into a typed database — points items, English thresholds, per-subclass criteria, occupation lists — each row linked to the exact source version it came from.

4. 4

   Verify

   Three automated gates must pass before a figure is trusted (below). This is what replaces a human proof-reader and makes “no uncited figure” an enforceable property, not a promise.

5. 5

   Register

   Record the source in the provenance registry with its authority level and as-at date. Nothing is “known” until it is registered against a saved primary source.

6. 6

   Index

   Make the corpus queryable — keyword/full-text search over the legislation plus direct structured lookups the assessment engine calls.

7. 7

   Detect change

   Watch each source's version endpoint. When the law is amended, re-fetch, re-verify, and flag every downstream figure (and any past answer) that cited the changed provision.

Why you can trust a figure

Stage 4 is the heart of it. Three gates run automatically; a figure is only trusted if all three pass.

If the knowledge base can't verify something — for example the full grant criteria for a subclass — ozimmi declines rather than guesses, and flags it for a registered professional.

How coverage extends across pathways

Australia has 100+ visa subclasses, but they are not 100 separate documents. The grant criteria for each subclass live in one instrument — Schedule 2 of the Migration Regulations 1994, one Part per subclass (the schedule runs across a few compilation volumes). So extending coverage is a parsing job over a handful of documents, not a hundred manual efforts. The ingestion order, by leverage:

1. Migration Act core sections (validity, grant, character, the points machinery).
2. Schedule 2 criteria, split per subclass.
3. Public Interest Criteria (Schedule 4, e.g. the integrity criterion).
4. The occupation-list and regional instruments (these change most often — wired to change-detection first).
5. Points and English instruments (already ingested), then policy guidance and tribunal decisions.

A subclass only counts as fully ingested when it has all four building blocks — points, English, Schedule 2 criteria, and its occupation list. We publish that as a live coverage metric so breadth is measured, never merely claimed.

How currency is tracked

Australian migration law changes often, so currency is treated as a property, not a hope:

• Version-pinned. Every figure is tied to the exact compilation it came from and shown with an “as at” date.
• Currency gate. A source past its re-verify window is flagged stale and surfaced as such — never quietly served as if current.
• Change detection. Each source's version endpoint is watched; when the law is amended, the affected figures and any past answers that relied on them are flagged for re-verification.
• Auditable. Every answer records which source version it used, so any figure can be traced back to the byte it came from.