The Australian medicine subsidy record
An independent, grounded record of Australia's HTA decisions — the PBAC's published outcomes, the Public Summary Documents that explain them, and the PBS expenditure that follows. 4,271 decisions 2001–2026, with 3,023 public summary documents across 1,047 drugs. Built from public data on pbs.gov.au.
Those are two different counts and the difference matters. A Public Summary Document can cover more than one meeting, and not every decision produces one — we hold a document for about 60% of published decisions. Counts of decisions come from the PBAC's meeting outcomes; the documents are what explain them.
What you're looking at
The Pharmaceutical Benefits Advisory Committee (PBAC) decides which medicines the Australian Government subsidises through the PBS. After every meeting it publishes a Public Summary Document (PSD) for each application — what the sponsor submitted, what evidence was provided, and whether it was recommended for listing. script.report aggregates every PSD published since the public summary regime began, alongside the actual PBS expenditure once a medicine is listed.
Who it's for
Health economics, market access, and HTA teams — wherever they sit. Australia runs one of the world's most rigorous HTA systems, and most multinational sponsors file a PBAC submission at some point. Australian specialists use the record for depth; international teams use it for their Australian submissions.
How we use AI — and how we don't
We use AI to read, index, code, and organise the public record. Humans still own the judgment.
Every Public Summary Document is extracted into a structured, auditable table — fields you can check against the source on pbs.gov.au. Search and optional grounded answers sit on top of that structure. No claim is shown without a verbatim quote from a listed record; if the record cannot support a claim, we stay silent.
Formal recommendations and funding decisions remain a human responsibility. We work only with published public documents — not confidential submissions — and we do not treat a model output as a substitute for professional judgment.
Working exports are meant for the tools you already use (spreadsheets and documents). No particular AI product is required to use the record.
Claude Science and the MCP connector
The same structured record is available to AI assistants over a Model Context Protocol (MCP) connector. It is a dependency-free stdio server exposing nine read-only tools over the public PBAC record: precedent search, a drug's full decision history, PBS spend, the revealed willingness-to-pay picture, and the Field Matrix (browse, filter, and per-drug detail). Every response is computed from the public documents and cites script.report. Add it to Claude Code with one line:
claude mcp add script-report -- python -m script_report mcp
It runs entirely on the public record, so there is nothing sensitive to approve. Browse the same data on the Field Matrix.
Where the data comes from
| Decisions | The PBAC's own post-meeting outcomes summaries, published per meeting on pbs.gov.au — these enumerate what the committee decided, whether or not a document followed |
| Reasoning | Every PBAC Public Summary Document (PDF + HTML) — the comparator, economics and stated reasons behind a decision |
| Expenditure | Monthly PBS expenditure statistics from Services Australia |
| Schedule · ATC | The monthly PBS Schedule API |
| Meetings | PBAC cycle calendar PDFs and published meeting agendas |
Every figure on the site traces back to a public pbs.gov.au URL. Source links sit on each drug page.
How the dataset is built
PSDs are 4–80+ page documents written in dense committee prose; the fields that matter — indication, comparator, evidence type, ICER, recommendation, rejection reasons — are buried throughout. Each PSD is normalised into a structured row of extracted fields, surfaced on the drug page with a backlink to the source document. Extraction is done by a language model and checked against the source; it is good but not perfect (see limitations).
Precedent search
Every drug's decision profile is converted into an embedding vector, one per distinct drug-and-indication, and your query is matched against the whole corpus by similarity. PSDs share a tight technical vocabulary — comparators, ICER framings, ESC commentary, managed-entry mechanics — so similar precedents end up neighbours even when the drug names, indications, and wording differ. A query like "immature OS in second-line oncology rejected on uncertain ICER" returns the precedents the PBAC has actually weighed, not keyword hits.
The grounded answer above the results is held to a strict contract: every claim must be backed by a quote copied verbatim from a shown record, and claims that can't be are dropped. When the records can't support an answer, it says so rather than improvising one.
AI practices, grounding & standards
Where AI is used — and where it isn't. Language models assist in exactly two places: extracting each PSD into a structured 41-field schema, and composing the grounded answer in search. Everything else on this site is deterministic computation over those extractions. No AI writes editorial content, estimates figures, or fills gaps.
Extraction. A fixed prompt against a versioned schema (currently version 4); a deep read of each document's economic sections; schema bumps force selective re-extraction so every row records which version produced it. A confidence pass adjudicates calibration-critical fields with a second model; its coverage is partial and expanding. Prompts and schema are available on request for reproducibility.
The grounding contract. Every claim in a synthesised answer must carry a quote copied verbatim from a specific decision record. A deterministic function drops any claim whose quote isn't an exact substring of its source, whose drug wasn't in the result set, or that asserts an outcome the structured fields don't support. The contract is enforced in code and tested, offline and against the live site.
Measured, honestly. Retrieval quality is scored against a fixed labelled gold set: precedent recall@10 was 0.797 when the set was built (May 2026) and 0.762 re-measured in July after the corpus grew by roughly a third — the drop reflects new, unlabelled candidates entering the top ranks, and the gold set is due for re-anchoring. We publish the movement rather than the flattering number.
Standards. The approach is designed for compatibility with the ISPOR working group guidance on generative AI in health economics and outcomes research — the ELEVATE-GenAI reporting guidelines, the GenAI-for-HTA opportunities and policy report, and the GenAI taxonomy — model identity, versioning, human oversight, and grounding are documented and auditable.
Independence. Formal PBAC recommendations are made by people, not systems. This site is an independent research aid on the public PSD and PBS trail. It is not affiliated with the PBAC, the PBS, or the Department of Health. Where generative tools assist extraction or cited search answers, that use is documented on this page so it can be inspected.
Limitations
- Extracted, not calculated. Figures such as ICERs reflect what was discussed or referenced in committee documentation — not independently confirmed inputs. For authoritative cost-effectiveness data, refer to the original PSD.
- Extraction errors exist. Some fields will be wrong, especially in older or unusually formatted PSDs. Always trust the source PSD over the extracted field.
- Pricing is redacted in public PSDs. Published ICERs are nominal submitted values, not the final price negotiated through any managed-entry agreement.
- Coverage is the PBAC PSD record only. Non-PBAC subsidy programs (LSDP and others) are not included.
- Not advice. This is a research aid, not legal, clinical, or commercial advice.
Get in touch
script.report is built and maintained by Josh Ciardi — an independent project, not affiliated with the PBAC, PBS, or the Department of Health. Something looks wrong, a PSD is missing, or you want a data cut the dashboard doesn't do? Send a note.