Standards · Rev. 3 · Oct 2026
Method
How a curated knowledge base becomes a published argument — and where a reader can check our work.
We publish analysis, not summaries. A summary can be assembled from retrieval; an argument requires someone to decide what matters, defend the choice, and accept correction. This page documents how that decision is made, who signs it, and what we owe a reader when we are wrong.
Our knowledge bases are built before a question is asked. Each research desk maintains a curated corpus of primary documents — filings, standards, statutes, technical specifications, operator disclosures — indexed at the paragraph level and dated at the moment of entry. Retrieval at query time is not research; it is recall.
- Primary sources take precedence over commentary, and commentary takes precedence over news coverage.
- Every figure in a report carries a source identifier and a date of record, visible in the online version.
- Where sources conflict, the conflict is described in the text rather than resolved silently in a footnote.
- Anything we could not verify is labelled as such, or left out. It is not softened with hedging language.
- Machine-assisted retrieval is permitted; machine-written conclusions are not. Every claim has a named author.
Each material claim is tagged with one of three levels. The tag reflects the distance between the claim and the evidence, not our enthusiasm for it.
| Level | Meaning | Test a reader can apply |
|---|---|---|
| Observed | Stated in a primary document we can cite | Open the source, find the sentence |
| Derived | Arithmetic or inference from observed figures | Re-run the calculation |
| Estimated | Judgement where evidence is partial | Check the stated assumptions |
Estimates are legitimate. Presenting them as observations is not. When a report's central conclusion rests on an estimate, that is stated in the first two paragraphs rather than the last.
Three signatures are required before publication: the lead analyst who owns the argument, the data editor who owns the numbers, and a review-board member who was not involved in the draft. The reviewer's mandate is to attack the method, not to improve the prose.
Review is recorded. We keep a dated log of the questions raised, including those we chose not to act on, and we publish the reviewer's name in the report metadata. Anonymous review is appropriate for academic journals protecting junior authors; it is not appropriate for a publication asking to be trusted by strangers.
Corrections are published, dated and linked from the original report. We do not quietly edit an archived version; the archive keeps every revision.
- A reader or reviewer files a challenge with the specific sentence and source.
- The data editor re-checks the underlying record within five working days.
- If the claim fails, the report is revised, the change is logged, and the revision number increments.
- If the claim holds, the challenge and our response are added to the log — a correction that was not needed is still information.
Corrections log
No. 014 · Rev. 1.2 · 27 Oct 2026 — Figure 1 caption clarified to mark values as illustrative.
No. 011 · Rev. 1.1 · 21 Sep 2026 — Interconnection range corrected in §3; source date updated.
- No paid coverage. Research is funded by subscription and licensing. No report is commissioned by a subject of that report.
- Disclosure. Commercial relationships, including data licensing, are listed on every report page.
- AI use is declared. Where machine systems assisted retrieval, drafting or translation, the report states which stage they touched.
- No individual targets. We analyse institutions and systems; we do not publish claims about named private individuals.
Corrections, methodological challenges and licensing enquiries go to the same address, and are read by the data editor rather than a marketing team.
desk@globalinfinitylabs.com
Global Infinity Labs · Singapore
This page is part of a design prototype. Policy text is illustrative and must be reviewed by counsel before publication.