Data downloads

Everything on this site is generated from one dataset, and you can have it. These files are rebuilt with every deploy (this copy: 29 Aug 2026), so they always match the pages. No API key, no repository clone, no scraping required.

The files

statements.csv

One row per tracked agency (135 rows): roster, coverage status, currency, the Standard report card, and the current profile, flattened for a spreadsheet or dataframe. Start here.

statements.json

The same 121 statements with structure intact: the full current profile (including commitment texts) plus every revision's classification, model-written summary and field-level profile deltas.

timeline.json

Every observed change as a flat event list (469 events, newest first), each with its classification and summary.

adoption.json

The monthly adoption series behind the policy page's charts, plus every agency-level concept transition with the revision it happened in.

Reading the CSV

Columns fall into five groups. abbr through originality are the roster: agency name, portfolio, whether the policy binds it (scope), coverage status, statement URL, when we first saw it, how often it has changed, and its lexical originality score. stated_first_published through annual_review_overdue are currency: the statement’s own dates against the policy’s review rules. The standard_* columns are the report card — which of the Standard’s eight minimum elements the statement shows. The remaining columns are the current profile in the closed vocabularies of profiles.py, with list-valued fields (usage patterns, domains, measures, named tools) joined by semicolons and profile_model naming the model that read it. Agencies with no published statement keep their roster columns and leave the rest empty — a blank profile means “no statement”, not “answered no”.

What to know before you analyse

The profile fields, change classifications and summaries are a language model’s reading of each statement — convenient, mostly right, and worth spot-checking against the linked text before you publish a finding. Reading the data explains what each figure claims and, more importantly, what it does not. The record also only starts when the tracker did: a statement’s first tracked revision is not its first published version.

The schemas follow the extraction code and do evolve. Each JSON file carries generatedAt, headSha and schemaVersions so you can tell which vintage produced your copy; new CSV columns are appended rather than reordered.

What is not here

The statement texts themselves. Full markdown bodies for every revision live in the git repository, and the current corpus is mirrored to the AT Protocol network under @apsaitracker.app as one document record per statement. The passage-reuse clusters behind the propagation explorer are also derivable from those texts.

Reuse and citation

The dataset — these downloads and the derived fields in them — is licensed CC BY 4.0, the same licence most Australian Government open data carries: use it freely with attribution. The scraper and this site are MIT-licensed, and the statement texts remain the work of their agencies, most themselves licensed CC BY 4.0. If the data feeds something published, cite the tracker via the repository’s CITATION.cff, and check the about page for the fuller licensing story.