Standards
Editorial standards
What we publish, how we check it, and what is not built yet.
Hierarchy of evidence
- Statutory text or a commenced provision, for a claim that a rule is law.
- HM Treasury or HMRC final published policy or technical notes, for what government has announced or proposes.
- OBR publications, for what the independent forecaster assesses or forecasts.
- ONS statistical releases, for historical observations.
- Commons Library and other official independent analysis, for context.
We never turn a policy announcement into a statement of current law, and we distinguish UK-wide from devolved Scottish applicability, tax years, rounding and nominal from real values.
Methodology
- Discover. Official collection pages, the GOV.UK Content API and OBR pages are polled at conservative intervals. A changed index page is a signal, never a story.
- Capture. The exact bytes are stored with the URL, redirects, timestamps and SHA-256. A new version is recorded only when content genuinely changes.
- Extract. Text is read from the document with page, sheet or cell coordinates preserved.
- Check. Deterministic release gates test the source and the stored copy before anything can be published.
- Decide. A human approves, or, only if the owner has enabled it, the narrow factual-notice policy applies. Human review is never silently bypassed.
What is built and what is planned
This table is kept in line with the project’s own status ledger.
| Control | Status |
|---|---|
| Allow-listed source polling with conditional requests, rate limits and backoff | Implemented and unit-tested; running |
| Immutable capture: URL, final URL, timestamps, SHA-256, version history | Implemented and unit-tested; running |
| Text extraction from HTML, JSON, PDF, CSV and XLSX with page / sheet / cell coordinates | Implemented and unit-tested (not run on every document type yet) |
| Deterministic release gates for document notices (host, reachability, stored-copy hash, timestamp window, duplicates) | Implemented and unit-tested |
| Human approval, kill switches, correction records with history | Implemented and unit-tested |
| Automatic publication of the narrow factual notice class | Built, switched off pending owner approval of the policy |
| Claim extraction and numerical checks (units, years, jurisdiction) | Planned, not built |
| Independent verification by a second model family | Planned, not built |
| Analytical stories (forecast vs policy, tax effects) | Planned, not built; would always need human review |
| LinkedIn distribution | Not enabled; no organisation posting permission yet |
Every material number is traceable
When analytical stories exist, numeric claims will be recorded with the exact quote or table cell, original URL, page or sheet, unit, reference year, jurisdiction and source timestamp. Claims without accessible evidence stay in draft.
Release gate
- All sources are official and available at the time of publication.
- Policy status (proposal, announcement or law), dates and jurisdictions are explicit, or stated as not assessed.
- Duplicate and corrected-document detection.
- Where a check cannot be completed, the story waits for human review.
Stages
DISCOVERED → CAPTURED → EXTRACTED → CLAIMS_DRAFTED → VERIFIED → APPROVED → PUBLISHED → DISTRIBUTED
Any stage may be blocked, sent for human review or retracted. Publishing to this site and distributing elsewhere are separate steps. Publication can be paused with a manual kill switch.
AI disclosure
Monitoring, drafting and checking are done by AI systems. Stories are credited to the VERADIQ editorial system, not to a named person. AI Cloud Craft Ltd and its human director remain accountable.