CYBER-HOBO
concepts
in practice
how it is built

Keeping rules honest

A set of instructions is not free to grow. Compliance with any one of them falls as the number rises, which is measured behaviour in language models. So a working rule set has to be small, and staying small means rules get removed. This is the test each rule has to pass to stay in, and it is four questions.

The four declarations

Every rule in the set states these four things about itself, in the same place the rule lives. A rule that cannot state them is re-engineered until it can, or it is cut.

where it appliesthe work it binds. Scope is the first of the four that turns out to be vague when you try to write it down.
what enforces itthe thing that makes it fire. Tooling at a fixed moment, a person noticing, or a review on a date. Name which, because the three are not equally reliable and the difference matters more than the wording of the rule.
what it caughtone specific thing it stopped that would otherwise have gone out wrong. Not what it is for. What it caught.
what would kill itthe observation that means it should be removed, written now rather than argued later. A rule with no stated end has no way to end.
ANSWERS ALL FOUR ANSWERS THE EASY TWO a rule that declares where it applies the work it binds enforced by tooling, at a fixed moment what it caught one named thing, on record what would kill it a stated observation KEPT, AND CARRYING A REVIEW DATE a rule that cannot where it applies everywhere, broadly enforced by someone remembering what it caught what would kill it DECORATION, CUT
The same four questions, two answers. The blanks on the right are the verdict, and they are always the last two. Scope and intent are easy to write for any rule at all. A rule that cannot name what it caught or what would end it has no evidence that it does anything, and it is competing for attention with the rules that do.

What counts as a catch

The third question is the one that does the work, and it is easy to answer dishonestly. Two answers do not count.

The first is what the rule is for. Every rule has a purpose, including the ones that do nothing, so a purpose cannot separate them.

The second is how often it fired. A count of firings looks like evidence and is not. What matters is a log of catches, not a log of firings, because a mechanism that fires constantly and has never changed an outcome is reliable theatre. The bar is counterfactual: name the specific thing that would have gone out wrong had the rule not been there. If nothing can be named, the honest entry is that nothing can be named, and that entry is what the review acts on.

not a catch"it keeps the work consistent." That is the purpose. Every rule has one and none of them are falsifiable.
not a catch"it ran four hundred times." That is a firing count. It says the mechanism is alive, not that it is useful.
a catch"it refused a document that would have gone out with a missing source, on this date." One instance, named, with the outcome it prevented.

Removal is the normal outcome

The test only means something if rules are actually removed by it, and a set where nothing has ever been cut is a set where the test is not being run. Four shapes of removal recur, each with the evidence that justified it.

what the record showedwhat it means
fired regularly, caught nothing in its lifetimethe class of error it targets does not occur in this work, or something upstream already catches it. The mechanism is cost with no return.
caught one thing early, then went cold for monthsusually a real catch against a habit that has since changed. The rule was right once. That is not a reason to keep it now.
caught nothing the surrounding tooling was not already catchingredundant enforcement. Something upstream already does the job, so this mechanism is paying for a boundary that is already held.
never used in practice, whatever the intentthe honest case, and the most common. Something was built, it was reasonable, nobody reached for it. It goes.

Recording why each one went is part of the removal. A cut with its reason on record can be reversed on evidence. A cut with no recorded reason leaves nothing to stop the same mechanism being proposed again.

What the test does not do

It does not tell you whether a rule is a good rule. It tells you whether the rule is doing anything. Those are different questions and only the second one can be answered from a record, which is the whole reason the test is written this way.

It also cannot force the removal it recommends. A rule with a blank catch column is usually somebody's considered principle, and the declaration only helps if a blank is allowed to stand as a verdict rather than as a task for later. The test makes the state visible. Acting on it stays a decision.