About

About this dataset

Australian federal MPs and senators must declare their interests (shareholdings, trusts, property, directorships, gifts, sponsored travel, memberships and more) in the Registers of Members' and Senators' Interests. The registers are public but published as scanned or typed PDF statements, one per member per parliament. This dataset turns them into one row per declared item, with a link to the page each item came from.

The site describes what was declared. It does not estimate value or wealth, and it does not suggest that any declared interest is improper.

Method

  1. Collect. House statements are the PDFs on aph.gov.au (archived registers for the 43rd-47th parliaments, the live register for the 48th); source_url points at each one. Senate statements come from the Parliament's senators' interests API (JSON).
  2. Transcribe. Each House PDF (initial statement plus every alteration bound into it) was transcribed by google/gemini-3.8-flash into a strict JSON schema, one item per disclosed interest, with anthropic/claude-sonnet-5.5 taking any pages Gemini refused. On a 12-PDF hand-checked gold set (790 items) this scored precision 0.986, recall 0.987 (F1 0.987), with 100% owner and page accuracy. Senate items are mapped from the API's fields; no model is involved.
  3. Validate and load. Every transcription passes a schema and completeness check (every page covered, every item on a real page) before loading. Members are matched across parliaments and chambers to one member_id; party and bloc come from a per-term table.
  4. Standardise entities. Entity names are normalised, then resolved in order: a curated alias table (the most common names), the ASX listed-companies list, an LLM grouping of spelling variants, and finally one entity per remaining one-off name. 11,542 entities.

Accuracy

Gold-set scores for the House transcription (12 PDFs, 790 hand-checked items)
MeasureScore
Precision0.986
Recall0.987
F10.987
Owner accuracy100%
Page accuracy100%

Senate items come from the Senate register's structured data: no model is involved. Each item carries the transcriber's confidence; items marked medium or low confidence are flagged on every page.

Prompt versions

Two prompt versions. Most House statements for the 43rd-47th parliaments were transcribed with prompt v0. Prompt v1 adds one rule (C12): itemise lists bound in as attachments ("see attached"). Only the 7 statements found to have un-itemised attachments were re-run on v1 (coultonm_43p, nevillep_43p, coultonm_44p, huntg_44p, pynec_44p, coultonm_45p, odowdk45p); the whole 48th House register used v1. A few other v0 statements may still describe an attachment in one item instead of itemising it.

Known limitations

How to read the pages

Contact and code

Made by Kevin Rassool. Found an error? Open an issue in the repository with the member, the item and the page of the statement. The code, documentation and full build log are in the repository.

Changelog