Qualifying public data for reuse: five stages, three candidates, one pass
A pipeline that resolves whether a public dataset may be reused, and whether it holds the quantity a question actually needs. Both gates refuse to resolve ambiguity by optimism.
Verdict
ONE OF THREE CANDIDATES QUALIFIED, AND RIGHTS REJECTED NONE OF THEM
Three real candidates were examined end to end. One passed all five stages. The two rejections were both on content: the data did not hold the measured quantity.
That inverts the expectation the project started with, which was that licensing would be the binding constraint.
Five stages, each refusing to guess
The pipeline resolves a candidate in five independent steps, and any stage may stop it.
| Stage | Question | Refusal rule |
|---|---|---|
| Identity | What exactly is this artifact? | Pinned by SHA-256, size and retrieval date. A recheck records drift and never rewrites the pin. |
| Rights | May it be reused? | Claims are collected as a set with sources. Absence of a licence never resolves to permission. |
| Content | What is actually in it? | Every format yields the same inventory: columns, types, units, ranges, missingness. |
| Question | Does it answer the question asked? | Required fields must resolve inside one table. A match assembled across unrelated tables is rejected. |
| Dossier | What is the recommendation? | Gated on the four stages above; identity, rights chain and missingness travel with it. |
The rights stage is where most of the judgement sits. A licence claim carries a scope, so the question is never simply what licence appears somewhere on the page. It is which claims cover the artifact itself.
Publisher policy attached to an article is contextual: it can never grant reuse of a data file. A licence in a repository root, in file metadata, or on the record can.
Three outcomes the pipeline treats as unresolved rather than permitted
An unknown label defaults to marker, never to permission. That single rule produces three of the most common outcomes in practice.
| Applicable claims | Verdict | Why |
|---|---|---|
| None | Unclear | Public download is not reuse permission. There is nothing to quote, and the verdict is itself the finding. |
| Markers only | Unclear | A label such as other-open or a bare copyright notice is not a licence. |
| Permissive only | Reusable | The governing licence is quoted in the dossier. |
| Restrictive only | Not reusable | The blocking term is quoted. |
| Permissive and restrictive | Unclear | A chain conflict, with both claims named. It is never silently resolved. |
Non-commercial terms count as restrictive here. The portfolio republishes results derived from the data, and a non-commercial grant does not cover that.
The candidate that passed, and the two that did not
The passing source is a thermophysical property database for deep eutectic systems. Rights resolve to CC BY 4.0 across two claims with no disagreement. Seven tables, 1,085 rows. All three required fields resolve inside a single table, with a per-row uncertainty and a reference DOI.
| Candidate | Rights | Content | Question | Outcome |
|---|---|---|---|---|
| Deep eutectic property database | Reusable, CC BY 4.0 | 7 tables, 1,085 rows | Match | Shortlisted |
| Measurement-level DES database | Reusable, CC BY 4.0 | 38 tables, 19,804 rows | No match | Rejected on content |
| Biomass-derived renewable solvent | Reusable, CC BY 4.0 | Inventory extracted | No match | Rejected on content |
The second rejection is the more instructive one. It is a larger and better-curated database covering density, viscosity, conductivity, surface tension, refractive index and melting point. It is entirely permitted to use, and it does not carry heat capacity.
That is precisely the split the pipeline exists to find before anyone writes analysis against a source: permitted, well-made, and missing the quantity in question.
The question was written to fit the data, and the record says so
The original question asked about enzymatic conversion by solvent system and substrate loading. No candidate passed it. Biocatalysis records overwhelmingly publish raw instrument output, chromatogram traces and peak areas, or compiled figure exports, rather than tidy conversion-by-condition tables.
Rather than report an empty deliverable, a question was written against data confirmed to exist. Reversing that order is legitimate for demonstrating that the pipeline qualifies real sources. It is not a substitute for a study-driven question.
The practical consequence is stated wherever the shortlist appears: the rights verdicts carry over to another study, and the question match does not. Anyone reusing this shortlist must re-run the matching stage against their own specification.
An audit found the matcher could pass a question it had not answered
An independent audit run after the pipeline was declared complete found two defects in the question-matching stage.
The first was structural. The matcher built one global registry of every column across every table, then searched it with no requirement that the matched fields share a table, a join key, or a compatible grain. Two unrelated files could therefore satisfy a three-field question between them.
The second was narrower and had the same character. Type checking was guarded by a lookup that had no entry for scalar requirements, so every numeric declaration on a scalar field was decorative. Text passed where a number was required.
Both are fixed, and the passing candidate was re-qualified under the corrected matcher. It still passes, now with all three requirements resolving inside one table.
The distinction matters more than the outcome. The verdict had been substantively supportable before the fix, but it had been established by hand afterwards rather than by the pipeline. A qualification that needs prose to excuse a matcher defect is not a qualification.
Provenance and reproduction
| Retrospective validation | Nine candidates with independently established verdicts, reproduced 9 of 9 from artifacts alone |
| Qualifying artifact | 116,937 bytes, SHA-256 c9e0d2421bc211d5175a6b4553b1140e6cd43620596ae72171ee10041d004f32 |
| Rights basis | CC BY 4.0, two applicable claims, zero disagreements |
| Required fields | Heat capacity, temperature and hydrogen-bond donor, all in one table of 110 rows |
| Matcher soundness | Pinned by split-table and text-where-numeric cases that fail against the pre-fix matcher |
| Acceptance gate | Deliberately fed a wrong expectation, and confirmed to go red |
The retrospective set is the acceptance criterion, and its pass is not degenerate. Every unclear verdict carries a recorded disagreement, and the three reusable verdicts resolve 3, 7 and 7 separate claims across landing page, file metadata and publisher policy.
That set earned its place once: a verdict defect in which a permissive claim silently outranked an unresolved deferral passed 178 unit tests and was caught only because the independent verdicts disagreed.
What this means
For well-run repositories, licensing is increasingly clean and the binding constraint is whether the dataset holds the measured quantity. All three candidates here were reusable; both rejections were on content.
A pass rate of one in three is consistent with the low rate this work predicted, and a high pass rate would be evidence of a lenient stage rather than of better candidates.
One limitation is unresolved and worth stating plainly. The test suite is not reproducible from a clean checkout, because the artifact cache and the confirmed retrospective records are not tracked. The reported pass count describes a host state rather than the delivered repository.