Ground rules: where facts come from and who decides

The places where you're most likely to guess are the ones you feel surest about.

That's why I keep a few written rules. They cover where facts come from, how facts turn into conclusions, and how conclusions turn into systems. Every model I use points back to these rules. None of them gets its own private version.

Evidence: if I can't find it, I write "not disclosed"

I only use information that can be checked. Every fact carries a source and a date, and a label for whether it's first-hand, third-party or unverified.

If I can't find something, I write "not disclosed" and leave the gap. I don't fill it with an inference. Early on I was tempted to. Then I realized the gap is a finding in itself. When a company doesn't publish a number, that often tells you something.

The rule that matters most is claim vs. fact. A user count the company reports about itself, its team size, the approval flow written into its policies, the requirements in a job post, how a candidate describes themselves. All of these go down as claims first, even when they're on the official site. They become facts once a real event or an independent source backs them up.

Having an approval policy doesn't mean authority is actually delegated. You find out by following one real decision and counting how many levels it went through.

One more thing. When two sources contradict each other, flag it and leave it. Don't smooth it over. Two pages of the same company saying different things is information. Marketing pages and contract terms that say opposite things are common too, and usually on purpose.

I rank sources roughly like this, from strongest to weakest: binding documents the company itself issued, audited or statutory records, job postings, material from agents and partners, signed media reports, interviews. In an interview, a specific event with a time, people and an outcome can count as a fact. Someone's opinion of themselves doesn't.

Judgment: every sentence hangs on an event

I almost never write "strong executor," "strategic thinker" or "good communicator." Those phrases carry very little information. Nobody reading them knows what to do next.

What I write is degree and condition: in which situation, what the person does. For example, "tightens control and stops delegating when pressure rises." That sentence is useful. You know what kind of seat to put them in, and what to watch for.

Every judgment has to hang on a specific event. If it can't, it's marked "to verify" and stays out of the conclusion. Where I can't see clearly, I leave it blank.

Blank is at least honest.

When I write the conclusion, I can only cite what's already been filled in above it. If two parts contradict each other, I list the contradiction separately. I don't average it away.

Design: every rule says when it stops applying

Every rule in a system, whether it's a grade, a bonus band or an allowance, has to point to the fact or judgment it responds to. If it can't, it doesn't go in.

Structure and numbers live apart. Market rates and fee levels change every six months or so. The skeleton of a system shouldn't change with them. So the method holds only the structure, the parameters are stored separately, and changing a parameter never touches the structure.

Every rule also says when it expires and how it's withdrawn. An allowance with no exit condition, once paid, is almost impossible to take back. I've seen that happen more than once.

AI doesn't make the hiring call

AI is very useful. It drafts candidate lists, sorts material, compares events, and copies the research I did on one company across ten more using the same template. It does all of that faster than a person. I use it every day.

A few things I don't hand over. Which companies make the final list. Whether to hire this person. The veto item in a judgment. Someone has to live with the consequences of those calls, so that someone should be the one making them.

I also have a habit. For every company AI copied the research to, I take a few facts and check them against the original source. What AI produces is neatly formatted with every source filled in. That's exactly why it looks so credible, and why an error can ride along all the way to the end.

Every page after this one uses these rules. They first come into play when you draw the two maps. The method overview has the full order.

Shorter thoughts go on LinkedIn.