The short version
What this evidence tells us.
Evidence cut-off: 3 October 2026, with the APLNG extension checked 4 October. The manifest records 97 attempted source resources, of which 90 originals were archived (104.9 MB). Resources include release pages, methodology, reports and workbooks; they are not all independently verified datasets. Seven archive routes failed; one of those, NT Healthy Homes, was read through the web reader and is recorded separately.
The sources are official statistical/regulatory publications, company disclosures and relevant First Nations-led organisations. Company claims remain attributed. The selection is a purposive public-source research baseline, not a systematic review with a registered search protocol or exhaustive retrieval of every publication.
Figures retain their own dates, populations and scope. The complete evidence and original source links remain below.
Evidence and context
What was checked
extract.py checks every downloaded original and HTTP-header SHA-256, 82 selected source controls, the identities of 53 selected HTML tables, the ATO/APLNG exact-entity reconciliation, twelve months of country diesel inputs and five explicit calculations. Other narrative findings were checked against the cited pages; the script does not verify every sentence or independently validate source producers. Selected PDF tables were rendered and visually inspected: APLNG tax pages 5–6, aged care page 19 and NHSAC page 127.
observations.json contains typed values, observation period, population, geography, units, accounting basis, uncertainty note, exact source locator/URL and original hash. selected-tables.json preserves whole selected published tables as labelled strings; blanks, suppression flags and uncertainty columns are not silently converted to zeros or estimates. source-notes.json records publishers, visible release-date labels and unresolved publication metadata. All seventeen CTG workbooks are archived; selected land/sea/language cells are controlled, not every cell in every workbook.
Originals and full reading copies are local, git-ignored study material. Their collection does not grant permission to redistribute company reports, commercial property data or other third-party material. Source licensing and any unpublished statistical uncertainty still need review for a specific app destination. This is why no record is yet labelled ready for app adoption.
Evidence and context
APLNG money trail extension
Batch 09 adds Origin’s issuer-authored FY2025 and FY2026 results lodged with SGX. verify_aplng.py reads the three pinned PDFs directly and checks 80 row controls, both cash reconciliations, the original FY2025 comparative, sales volumes, profit, debt and the FY2025 tax calculation. Run python3 verify_aplng.py for a read-only check of aplng-money-trail.json. --write regenerates that derived file; it does not change originals. The earlier 82 controls remain a separately counted baseline.
The annual reports close the dated project-debt and cash-distribution gaps left by failed Origin landing-page retrievals. They do not establish every lender identity or net public return. The final FY2026 cash-tax disclosure of A$1,414m differs from the preliminary A$1,410m; both are retained with their source boundaries. Four relevant FY2026 financial table pages were visually inspected.
Evidence and context
Reproduce without network access
From this directory, run python3 extract.py --check using the repository's research dependencies (lxml, openpyxl, pypdf). It expects the existing money-flow dossier at ../australia-money-flows; use --money-dossier PATH if located elsewhere. A failure stops the verification. It does not refresh sources or overwrite originals.
collect.py reuses scripts/research/source_library.py from the repository, rather than introducing another HTTP client. Its numbered batch files enumerate requested public URLs. Restoration is not automatic: a tracked checkout without the ignored originals will fail verification. For a new retrieval, copy collect.py and the numbered batches into a separate empty directory without this manifest, then run python3 collect.py batch-01.json --repository /path/to/repository and subsequent batches there. Compare the new body hashes with the pinned manifest; HTTP headers and live pages may differ. Keep the new capture separate and review revisions before adoption. Do not overwrite the pinned manifest and call that the same verification. Already-archived originals are checked and never overwritten by the collector.
Successful retries replace that source's previous failed status; the manifest is a final-resource record, not a complete HTTP-attempt history. In particular the PC assistance landing page required its canonical trailing-slash URL. The original failed route remains in its batch request.
Evidence and context
Source catalogue
“Facts/tables” counts only the selected reproducible controls. A zero does not imply the source was useless: it may supply methodology, qualitative findings or an archived lead. It also does not mean every source has been fully read.
Additional source records: origin-annual-2025-sgx and origin-annual-2026-sgx are pinned in manifest.json; their full locators and extracted controls are in aplng-money-trail.json.