How we verify
Scanned and filtered by AI. Stress-tested and judged by humans.
Every Signal Scan, Change Tracker and Decision Intelligence briefing on this site goes through the same steps before it is published. AI does the reading at scale. Automated checks test the result against the sources. People decide what it covers and whether it goes out. If a check fails, the deliverable does not publish until the problem is fixed, or until the point has been re-checked another way and the reason recorded.
The six steps below follow The Decision-Grade Intelligence Method. The labels show who does the work at each step.
AI reads open, dated sources every week: regulators, statistics offices, central banks, think-tanks, researchers and quality journalism. Sources in other languages are read in the original, not through a translation.
Each source is graded from Tier 1 (governments, regulators and intergovernmental bodies) to Tier 4 (vendors and campaigners), must fall inside a set date window, and has its publication date read off the page itself, with the evidence kept. Tier 4 sources never carry a finding on their own.
Automated checks test every figure, quote, date and link against the sources (listed below), and a separate AI model tests whether each section argues what its sources argue. Each deliverable must also set out the strongest counter-argument, drawn from sourced evidence, and an analyst reviews the draft before it can be published.
An analyst sets the scope of each topic, signs off the themes, title and readers of each Decision Intelligence briefing before the full research starts, and checks every deliverable before it is published.
At the end of each cycle, Action Triggers reads across everything published on a topic and sets out what to decide, prepare for or monitor, each with an owner and a date. An analyst confirms which deliverables it draws on and when it is published.
Every deliverable is published with its sources linked and dated. Decision Intelligence briefings add a full source register, an appendix saying where the analysis goes beyond what the sources say, and a ledger grading the calls the previous edition made.
These run on every Signal Scan, Change Tracker and Decision Intelligence briefing. Action Triggers and the Live Summaries draw only on deliverables that have already passed them, and every figure on a Live Summary is traced back to one of those deliverables.
Every source's publication date is taken from the page itself and the snippet it came from is stored. A guessed or missing date stops publication.
Every link in the text must lead to a source in the register. A link typed by hand, or one that has drifted to another page, stops publication.
The publisher we name has to match the website the source sits on.
Every number in the text is matched to the source's own wording, kept on file. A figure with no source behind it stops publication.
At publication every direct quote is fetched again from the live page and must match exactly. A quote stitched together from separate passages fails.
A separate AI model, which did not write the deliverable, compares what each section argues with what its sources argue. A section that bends a source to fit the story fails.
Where the analysis goes beyond what a source says, the deliverable says so in its claim-fidelity appendix.
The text is checked for jargon and stock phrases, so readers outside the sector can follow it.
Each Signal Scan states how likely its signal is to bite, as a band with a percentage, within a dated window of no more than three years, so a later edition can say whether it did.
Where a briefing gives the size of a market, the figure must come from a statistics office, regulator or comparable research body, never from a commercial market-sizing firm.