Editorial policy: sourcing, and where AI stops
How topics get chosen, how facts get checked, what AI does and does not do here, and what happens when we get something wrong.
How topics get chosen
One test: is anyone actually searching for this, and are the existing answers good enough.
In practice that means starting from real tasks: what do I do now my coin has been delisted, which chain should I withdraw on, can I still get funds out of a restricted account, rather than from concepts. Which is why there is nothing here explaining what a blockchain is: that has been written well many times and another version adds nothing.
Where we find we have nothing more to say than what already exists, we do not write it. Better to publish fewer pieces than one that restates someone else's point in different words.
How facts get checked
Facts on this site fall into three categories, handled differently:
| Type | Example | How it is handled |
|---|---|---|
| Protocol constants | Bitcoin averages a block every ten minutes | Written directly; these do not vary by platform |
| Platform rules and rates | A fee tier, a minimum withdrawal, a verification level | Written as a range and a direction with the month checked, plus a pointer to the current figure |
| Things that change with events | Whether a region is open, whether a coin is listed | No lists published — only how to check, because a list is guaranteed to go stale |
The second category is where errors are easiest, so we hold a hard rule: anything a platform can change at will does not get written as a fixed number. When you see "in the 0.1% band (checked 2026-08)" rather than a bare percentage, that is why.
Where AI is used, and where it stops
Stated openly: AI assistance is used in producing this site's content.
The boundary is drawn like this:
AI is involved in
- Organising and drafting
- Language work, shortening sentences, replacing jargon with ordinary words
- Bulk consistency checks, such as finding broken links across the site
People decide
- What gets written and what does not
- Every factual claim, and whether it survives checking
- Everything involving risk or judgement
- Whether it publishes
We do not think "was AI used" is the interesting question. The interesting question is whether anybody is accountable for the result. So the rule is: however a draft came about, every sentence is read by a person before publication; anything involving a number, a rule or a risk is checked against a public source line by line; and whatever cannot be verified gets deleted rather than hedged with "approximately".
A related rule: no AI-generated illustrations. Images here are either drawn programmatically or are screenshots of real web pages, captioned with the month they were taken. AI images tend to carry traces that are hard to see, and using fabricated images on a site about how to verify things would be a poor look.
On the first person
Some passages say "what I do is" or "I basically never use this". Those are the writer's preferences and trade-offs, marked as such so you can tell which sentences are judgement and which are fact.
At the same time we do not do the following: invent test timings accurate to the minute, claim evidence we cannot produce, or fabricate trade amounts and returns. If a passage reads as though it rests on hard evidence, it should have some. Where we cannot produce it, the language becomes "usually" or "roughly" rather than a manufactured figure.
What happens when we get it wrong
-
Confirm it
Go back to the original source. If the platform changed a rule rather than us writing it wrongly, the handling differs: the first is an update, the second is a correction.
-
Fix the text
Edited in place, without strikethrough versions preserved, because that makes pages progressively harder to read.
-
Log it
What changed, when and why goes into corrections. The more substantive the change, the more detail it gets.
-
Propagate it
Where the same fact appears in several pieces, all of them change together. No stale copy is left behind.
Anything that would affect an operational decision goes through the full process; typos and formatting are fixed silently.
Update cadence
The content splits in two. The method parts — how to judge, what order to work in, are relatively stable. The parts touching platform rules and numbers need periodic re-checking, and we update those as platforms change, refreshing the date shown on the page.
If a page's date looks old and it concerns something changeable, defer to the exchange's current page: every such passage says where to look.
Tell us
Errors, stale material, anything unclear — contact details are here. Where a point stands up we will change it and log it.