How we assess

This register makes two kinds of claim: that a dataset is available in a certain form, and that a question can or cannot be answered from what is available. Both are falsifiable, and this page explains how we arrive at them so you can challenge either precisely.

We start from questions, not datasets

Every open data portal in Australia is organised around holdings — a list of what agencies chose to publish. That structure has one unavoidable property: it can never show you what is missing. A catalogue of what exists cannot contain the thing nobody holds.

So the atomic unit here is the question: something a person actually tried to find out. Datasets are attached to questions rather than the other way around, and a question with no dataset behind it is the most valuable record in the register.

Answerability

Each question gets one honest verdict:

Blockers, and the distinction the whole site turns on

Where the verdict is not yes, the question records why. This is the most important field in the schema, because "the government won't tell us" is not actionable and "it is published quarterly as a PDF with no consistent category definitions" is.

Blockers divide into two kinds, and keeping them apart is what makes this register useful rather than merely aggrieved:

Publishing choices — format, licence, aggregated, fragmented, inconsistent, stale, not_published. The data exists. Someone already collected it, at public expense. What stands in the way is a decision about how it is released.

Constraints on the world — not_collected, exempt, commercial, identifiable, paywalled. Either nobody holds the information, or releasing it raises a question that is not merely administrative.

A question blocked only by the first kind is flagged fixable without new collection, and those are the strongest cases in the register. The standard objection to any transparency request is that collecting the data would be expensive. For these, it already has been.

not_published sits on the publishing-choices side, and that placement is deliberate. The single test running through the split is whether answering the question would require anyone to start collecting something. Withholding a report that has already been written, by officials already paid to write it, does not. It is a decision about release, made by the same people and on the same grounds as the decision to release something as an unqueryable PDF, and treating the two differently would let the more complete refusal score better than the partial one.

That placement makes the other side carry more weight, so it is applied strictly. not_collected means nobody has the answer and someone would have to start counting. exempt means a statutory ground is claimed. commercial means a contract restricts disclosure of its own terms. identifiable means release would expose individuals. paywalled means access is real but priced. A question carrying any of these is not scored as a free fix, however strongly the register might feel about it.

commercial needs the same discipline identifiable gets, and for the same reason. Contractual confidentiality is a real interest: a government that could not promise it would pay more for everything it buys. Recording it honestly is what earns the register the standing to say, in the next sentence, that it rarely covers as much as is claimed for it. A confidentiality clause over a price schedule does not obviously extend to how many units were delivered, and the useful contribution — exactly as with identifiable — is to describe the version of the disclosure that would survive the constraint rather than to demand everything and be refused.

identifiable deserves particular care. Some data genuinely cannot be released at fine granularity without exposing individuals, and recording that as obstruction would be dishonest. Where it is the real constraint we say so, and the useful contribution is to describe what could safely be published instead.

The usability score

Datasets are scored out of 100 on how usable they actually are — not on whether they exist:

• up to 40 for how freely available it is (open, published, paywalled, on request, FOI-only, withheld, unpublished, not collected — in that order)
• 20 if it is machine-readable
• 15 if the licence permits reuse
• 10 if there is an API
• 10 if the granularity is recorded and meaningful
• 5 if an update frequency is stated

A PDF on a website is published. It is not usable. The score says so, and the gap between those two claims is where most of the real argument about transparency lives.

Anything unpublished, not_collected or withheld scores zero regardless of its other qualities.

Publisher scores: read the median

Publisher pages report both mean and median dataset score. The median is the honest number. A body that publishes one excellent flagship dataset alongside forty PDFs will have a respectable mean and a poor median, and the median is what a member of the public actually encounters.

Three things the scorecard cannot tell you, stated plainly because a score without its limits is a weapon:

1. Coverage is not completeness. A publisher is scored only on the datasets we have assessed. Two good datasets and a hundred unexamined holdings produces a good score and means nothing.
2. A high score is not a clean bill of health. A body can publish beautifully machine-readable data about trivia while withholding everything that matters. Read the gap register alongside the scorecard.
3. Small bodies are not badly run because they score low. A council of four thousand people has no data team. Comparisons are only meaningful between bodies of similar size and budget.

What we have not checked

Every dataset carries a last_checked date, and most are currently empty. Those entries are recorded from documentation rather than from opening the data ourselves. They are listed publicly on the verification page rather than quietly carried as if they were verified.

A transparency register that has not checked its own claims is making exactly the mistake it criticises. Saying so is not a disclaimer; it is the same standard we apply to everyone else.

Corrections

If data is published somewhere and we have recorded a gap, tell us — a false gap is worse than no entry at all, and it is the failure mode most likely to cost this project its credibility with the people who could act on it.

See contributing.