Methodology

Editorial and source-verification principles

• Primary sources first. Every record links to an official document or page. Secondary sources are used only when clearly labelled and never as the sole basis for a legal claim.
• No invention. If a date, penalty, obligation or outcome cannot be established from the source, the field is left empty and the confidence level is set to "unavailable". We never fill gaps with assumptions.
• Original summaries. Summaries are written in plain language with article or section references. We do not reproduce copyrighted legal commentary, paid database content or standards text such as ISO/IEC documents.
• Visible uncertainty. Proposed laws, differing application dates, pending amendments and unconfirmed facts are labelled as such in the record and on the page.
• Human verification. A record is marked verified only when a named reviewer who has published their declared interests has opened the official source and confirmed each field. Until then it shows "Pending review" while it waits in the review queue, or otherwise "Source-linked" (linked to the official source, not yet confirmed by a reviewer), with the date the source was last checked where there is one. Verified means a factual check against the official source; it is not a legal review and not legal advice. Each record type has a maximum age before re-check; see the verification policy for the rules and the current standing.
• Completeness is published, not claimed. Every record kind has fields it must carry to be checkable at all. Coverage counts how many carry them, the open queue names the ones that do not, and the corrections log shows what readers reported and what was decided, refusals included.
• Not legal advice. Content is informational. It is not a substitute for advice from qualified counsel in the relevant jurisdiction.

Source-of-truth hierarchy

1. Tier 1: Official primary source (government, legislature, regulator, court, standards body, intergovernmental)
2. Tier 2: Official consultation, guidance, enforcement, procurement or agency source
3. Tier 3: Reputable institutional secondary source (clearly labelled)
4. Tier 4: Community submission awaiting review

Data model and taxonomy

A jurisdiction holds policy instruments. An instrument has versions, sections, obligations, applicability rules, deadlines, enforcement events, procurement rules and source documents. Obligations carry evidence examples and original framework mappings. Change events record dated developments. Controlled vocabularies cover ai actor, ai system type, sector, risk category, evidence type, ai use case, obligation category. The full schema is documented on the open data page and in the repository.

Review levels

Confidence levels (high, medium, low, unavailable) describe how directly the record is supported by the identified source, independently of whether a human has yet verified it.

Independent checks

A verification is only as good as the person who made it, so the site publishes how often two reviewers agree. Each quarter a random 20% of verified policy records is drawn with a published seed, so anyone can re-run the draw and confirm it was not hand-picked. A second reviewer from the roster, never the person who verified the record, re-reads the official source without seeing the first reviewer's values, records the status, whether the instrument is binding and whether it should stand as verified, and says whether they agree on the dates, who it applies to, the obligations, the penalties and the source. Disagreements are resolved between the two reviewers and logged on the record; a resolved error is corrected and appears in the corrections log.

No independent second checks have been published yet. Every verified record so far has been confirmed by one reviewer, whose name is on the record. Agreement figures will appear here once at least 20 records have been double-checked.

Policy status definitions

How changes are detected and reviewed

Maintainers and contributors monitor official journals, legislature and regulator websites, consultation portals and government press releases for the covered jurisdictions. Each development is recorded as a change event with its date, source, practical impact and impact level (urgent, high, routine), then reviewed before publication. Records affected by a change are re-verified and their content version incremented. The change log and its RSS feed are the public record of this process; the repository history shows every edit. The target is to log a change within 3 days of the official text; the freshness page measures how we are doing against it.

How "top stories" are ranked

The updates pages rank changes by a significance score from 0 to 100, computed by a published rule rather than chosen by an editor. It adds: the impact level (urgent 40, high 25, routine 8); 15 if the instrument is binding; 12 if the change is an entry into force, adoption or enforcement action, 4 if it is a proposal or consultation; 8 if a named reviewer has confirmed the record; 5 if it carries an official source; 5 if the jurisdiction is one the site follows closely; and a recency term (10 within a week, 5 within a month, −5 after a year). The score orders the list; each top story is labelled with its recorded impact level rather than the raw number. The rule is versioned in the repository (App\Services\Changes\Significance) so a change to it is visible.

Independence and permanence

Who verifies the records, the commercial relationship and the rules that keep it away from the data are set out on editorial independence. Every record has a permanent id (AIPT-POL-0073, AIPT-OBL-0004 and so on) that never changes or gets reused; /id/{id} always resolves to the record's current page, so cite the id.

Known limitations

• Coverage is deliberately narrow and deep: a small number of jurisdictions with structured records rather than a broad but shallow list.
• The coverage report measures completeness of the records that exist here, not coverage of the world: it cannot tell you how many instruments exist that have never been recorded.
• Many records are still pending human verification; the verification line on each page tells you which.
• Some official documents are available only in local languages or without stable URLs; such records say so in their source reference.
• Framework mappings to ISO/IEC 42001 and the NIST AI RMF are original editorial judgements with stated confidence, not official crosswalks.
• Application dates can change after enactment; a record is only as current as its last verification.

Disclosure of AI assistance

AI tools may be used to help draft summaries, structure records and detect changes. Every published record must still link to an official source and pass human review before it is marked verified; AI output is never published as verified fact on its own. The initial dataset was drafted with AI assistance and is labelled "pending review" until reviewers complete verification.