How it works

Every Australian state and territory maintains a public website where coroners' findings are published after they are finalised. An automated scraper runs weekly, visiting each of these official websites and checking for new findings. When a new finding is detected, the scraper downloads a copy of the PDF document and records the original source URL so users can always verify against the official version.

Only publicly accessible findings are collected — nothing behind a login, paywall, or court restriction. The scraper is rate-limited, pausing between requests so as not to overload government servers.

Some older findings are no longer on the courts' current websites, often after a site redesign. Where a public copy was preserved by the Internet Archive's Wayback Machine, it has been imported from there, and the archived address is recorded as its source.

Coroners' findings are published as PDF documents. These vary enormously in quality: some are modern, digitally typeset documents; others are scanned images of typed or even handwritten pages from older proceedings.

The pipeline first attempts to extract text directly from the PDF using a digital extraction library. For most modern documents this works perfectly. However, if the extracted text is too short or appears garbled — a sign that the PDF is a scanned image with no embedded text layer — the pipeline automatically falls back to Optical Character Recognition (OCR). OCR analyses the visual image of each page and converts it to machine-readable text, much like how a phone camera can read a printed menu.

The extracted text is cached so that it does not need to be re-processed on every pipeline run. It is shown under “Full text” on each case page, and an archived copy of the PDF is linked alongside a link to the official source, so you can always read the document in full.

Some court PDFs are “redacted” with a black box drawn over text that is still present in the file and can be copied out. Every PDF is checked for this before its text is extracted. When it is found, a properly redacted copy is made with the hidden text removed, and that copy is used for text extraction and served as the archived copy. The full text of those findings is also withheld from case pages.

Once the text of a finding has been extracted, it is sent to a large language model (Claude, made by Anthropic) with a detailed prompt asking it to read the finding and extract specific structured information. This is the core of what makes the database searchable and filterable.

The AI reads each finding in full — which can be anywhere from a few pages to over a hundred — and returns a structured summary covering the fields below. (For the very longest findings, the middle of the document is omitted and the AI reads the opening and closing sections, which set out the circumstances, findings and recommendations.) This step is analogous to a trained researcher reading the document and filling out a standardised case report form.

Coroners' findings frequently name the treating clinicians involved in a case — doctors, nurses, paramedics, and others. While these findings are public records, republishing full clinician names in a searchable database creates a meaningful privacy risk: a clinician's name could become permanently and prominently associated with an adverse outcome in a way that goes beyond what the original court record warrants.

To address this, the pipeline automatically detects doctor name patterns in the AI-generated summary, contributing factors, recommendations, and other text fields, and replaces them with an initial. For example, “Dr James Smith” becomes “Dr S.” The substitution preserves clinical context while removing the identifying information.

This applies only to the AI-generated fields. The full text and the archived PDF reproduce the public court document as published, so names that appear there are not altered (apart from the removal of text hidden under defective redactions, described in Step 2).

Medical terminology is not consistent. The same term can appear with different capitalisation, hyphenation, or spelling — British and American variants are common, such as “anaemia” vs “anemia” and “haemorrhage” vs “hemorrhage”.

A vocabulary standardisation step runs over each new case. It checks every clinical condition and procedure tag against a curated master vocabulary list and, where a tag is a variant of an existing entry, replaces it with the canonical form (Australian English spellings throughout). Terms that do not match any existing entry are kept as written and added to a review queue for potential addition to the vocabulary.

Drug names are normalised to their generic names, so brand names and older Australian names map to the same entry — for example, “Ventolin” becomes “salbutamol” and “frusemide” becomes “furosemide”.

Hospital names go through a similar process. A fuzzy-matching step compares each hospital name extracted by the AI against a master list of known Australian hospitals and snaps it to the canonical name if the match is close enough (for example, correcting “The Royal Melbourne Hospital” vs “Royal Melbourne Hospital”).

Search queries run against a full-text index built over all structured fields — including the AI summary, specialty, drug names, clinical conditions, procedures, contributing factors, and recommendations. The underlying search engine is SQLite's FTS5, which uses the BM25 ranking algorithm to sort results by relevance when a query is present.

Filters (state, specialty, error type, etc.) work as exact-match refinements on top of the full-text search. You can combine a text query with multiple filters — for example, searching for “insulin” and filtering to ICU cases in Victoria.

Note that the full text of the original PDF documents is not part of the search index. Search runs on the structured, AI-extracted fields only. This is an intentional design choice: full-document search tends to surface findings that mention a condition only tangentially, producing less relevant results for clinical learning purposes. See the Research page for more detail on this.

• Coverage is not complete. Some findings are not published online by courts. Some older findings may not have been captured. The database should be used as a learning and discovery tool, not as an exhaustive archive.
• AI tagging contains errors. Specialty classification, error typing, and drug identification are all AI-generated and imperfect. Treat them as a useful approximation, not a verified classification.
• Findings are legal, not clinical, documents. A coroner's finding reflects the coronial process and the evidence presented in those proceedings. Clinical details may be incomplete, stated at a lay level, or filtered through expert testimony.
• Context matters. No summary captures the full nuance of a multi-day inquest. Always read the original document before drawing any clinical or educational conclusions from a case.
• Redaction and non-publication orders are respected. All efforts are made to honour court orders suppressing identifying information. Some court PDFs have technically defective redaction — visually obscured by a black box but with the underlying text still present and copyable. For those findings, the hidden text is removed from the archived copy before it is published here and never reaches the AI, and the full text is withheld from the case page. A finding is removed from the database if the court asks.

The scrapers run once a week. New findings published by coronial courts are typically added within a week of appearing on the official court websites. Because findings must be scraped, extracted, processed by the AI, and then indexed, there is always some lag between a finding's publication and its appearance here.

If you are looking for a very recent finding that you believe has been published, check the relevant court's website directly — they are all listed in Step 1 above. Each case page also links to the place the finding was published, under “Official source”.