In The Stamp of Legitimacy I ended on one line: watch the number, not the essay. A few readers asked what that means in practice. This is the short version.
A safety proposal has two parts. The first is the argument, which is written to persuade you. The second is a definition of who the rule covers. The definition is usually a number, and it is usually in a footnote or a scope section that few people read.
Check 1: Is the line a single test or a pair?
Anthropic's June 2026 framework is the clearest example I have found. It covers a developer only if that developer meets both of two criteria. The first is building models that need more than 10^25 training FLOP, a measure of training compute. The second is either more than $500 million a year in AI revenue or more than $1 billion a year spent on AI research and development.
Several news summaries described those as alternatives. They are not. The compute test is required, and then one of the two money tests must also be met. If you only read the coverage, you would get the scope wrong, and the scope is the whole point.
Check 2: Does the document say the line will move?
The same framework says the compute figure may need to change. In its own words, the training compute needed to build dangerous models may fall, and it may make sense to use a threshold based on capabilities instead of training cost.
That is an honest sentence, and it matters. A threshold set today is a snapshot. The question is what moves it, who decides, and which direction. A line that can only move down will eventually reach companies that were never in the room.
Check 3: Who falls just under the line, and what do they do?
The costs here are well documented in one place. A study by the Centre for European Policy Studies (CEPS) of the proposed Act's costs puts the cost of setting up a quality management system from scratch at 193,000 to 330,000 euros, plus about 71,400 euros a year to run it. Those costs apply to providers of high-risk systems that do not already have such a system. For anyone outside that class, they do not apply at all.
So the real line is not the cost. It is the definition of high-risk, because that definition decides who pays.
Then look at what people do near a line. A 2022 survey from appliedAI asked venture investors how the draft EU AI Act would change their plans. Nine of fifteen said they would focus on low-risk AI systems. Two of fifteen said they would back high-risk ones. That is a small sample taken before the final law passed, so I treat it as a signal and not a measurement. But it shows the pattern: money moves toward the side of the line with the lower bill.
How to do this yourself
- Find the scope or definitions section before you read the argument.
- Write down the number and say out loud who it includes and who it leaves out.
- If there are two tests, check whether both must be met or only one.
- Look for a sentence about how the number will change.
- Ask who is standing just below it.
None of this tells you whether a proposal is good. It tells you what the proposal does, which is the question that comes before whether it is good.
The research behind this post was done with Claude, which is built by Anthropic, one of the companies named above. I wrote about that conflict in The Bias It Could Name. The framework text, the cost figures and the survey result are all published, so check them yourself.
This is a note about how I work, not advice about how you should. I build software at Revelations Technology and write these as I go.