A living atlas of contesting and curating AI sovereignties (Australian view)

A snapshot of a live tracker of the physical infrastructure behind AI in Australia: data centres, mines, refineries, and the energy and water they draw. Each site carries its ownership chain, its approval pathway, and the public contestation forming around it. The intention is to evidence what version of AI sovereignty is being historically-produced in Australia. The live version is at https://datacentres.civicinterplay.io Data and methodology Version 0.1.0 was an initial snapshot of the project's code repository. From v0.2.0: the dataset, the datasheet, the method statement, the verification protocol, and the record of which claims survived adversarial checking and which were refuted and corrected. What's included sites.csv — 126 tracked sites and projects, one row each contestation_items.csv — 32 source events (articles, submissions, motions), joined to sites data-dictionary.md — every column defined, including what a blank means README.md — the datasheet: coverage, caveats, and how to reuse it METHODOLOGY.md — how the atlas is compiled, verified and corrected FACT-CHECKING-GUIDE.md — the working verification protocol VERIFICATION-RECORD.md — adversarial verification results, including the claims that were refuted DISCLOSURE-AUDIT.md — a completed audit of what Victorian planning records disclose about energy and water CHANGELOG.md — what changed between deposited versions Notes Throughout this dataset a blank field means the question has not been assessed; an explicit "Unknown" or "None" means it was assessed and nothing was found. The distinction is deliberate and it is what makes an absence citable. capacity_mw is the clearest case: most tracked sites have no capacity figure in any public record, and that absence is itself the finding. resource_conditions records whether the legal instrument of approval imposes any obligation on energy or water use, graded only from the instrument itself — never from an impact statement, an assessment report or a media release. governance_flags records findings about how an approval was handled; an absent flag means the row was not assessed on that dimension, not that it passed. How much of it is verified Claim-level. 28 claims have been put through adversarial verification, in which independent agents are tasked with refuting each claim from fresh sources: 20 confirmed, 2 plausible, 6 refuted. The refutation rate is 21%. All six refutations were upheld on independent re-check and corrected in the dataset and in published documents the same day. Every claim, its verdict and its failure mode are listed in VERIFICATION-RECORD.md. Separately, the Victorian disclosure audit checked each checkable planning record in its cohort against the exhibited primary documents; that audit and its caveats are in DISCLOSURE-AUDIT.md. Row-level. The classified_by column is a stricter and much narrower measure: it marks a row Human-verified only when a person has walked the row as a whole to a primary or official source under the fact-checking guide. On that bar, of 126 rows one is Human-verified, 45 are Human-entered, 21 are Agent, and 59 carry no classification. Rows classified Agent are unreviewed and are excluded from published figures. Claim-level checking has covered considerably more ground than the row-level flag records, because verifying one field on a row does not verify the row. Use classified_by and confidence to filter to the level of assurance your use requires, and read VERIFICATION-RECORD.md for what has actually been tested. Use of AI, on the record This project uses AI agents openly as part of an emerging method called her 'agentic civics'. Agents retrieve and search planning documents, draft row entries, and run research sweeps that must return evidence with URLs. Before findings are published, independent agents are tasked with refuting each claim from fresh sources, and claims are graded confirmed, plausible, refuted or unverifiable. Refuted claims are corrected everywhere they appear. Of 28 claims put through that process, 20 were confirmed and 6 refuted — a refutation rate we publish rather than hide, because it is the honest measure of what the method catches. Compute use is logged in the repository. Changes since v0.1.0 The full list is in CHANGELOG.md. The substantive ones: A governance-flag vocabulary was introduced and then reduced from 11 terms to 3. Flags that merely restated the statutory approval route were retired once the planning pathway field carried that information on its own. A First Nations engagement flag was tested and removed. Its six uses were concentrated in Western Australia and the Northern Territory and on mining projects — a record of where an analyst looked rather than where engagement is unclear. Because a flag is silent both when a row passes and when nobody checked it, an empty result could be read as an absence of concern. Every site in this dataset sits on Country. The replacement, when the work is done, is the statutory record — whether a Cultural Heritage Management Plan was required and approved, assessed with the Registered Aboriginal Party for that Country — not our judgement. The planning pathway vocabulary was made state-neutral. "State significant development" was previously used but is New South Wales statute, not other states. It now reads "State assessed", covering the equivalent route in each jurisdiction, with the local instrument named per row. A new field, resource_conditions, records whether an approval instrument binds energy or water use at all. Citing Cite the concept DOI 10.5281/zenodo.21026429 to resolve to the latest version, or this version's DOI to cite this snapshot specifically.