ContentsThe library

What a Model Refuses, and Why

If Two People Cannot Apply It, It Is Not a Rule

Last timeThe Price of Saying No

A safety intention becomes a rule when two readers can apply it to the same request and agree. Here is the rewrite that gets you there, and what the rest of the document is doing.

Seven lessons have been about mechanism: where a rule can live, what each

placement holds against, and what it costs to check and to refuse. None of them

asked what the rules should say. That gets answered in a document nobody reads

twice, and the document has one failure mode worth naming, because it is nearly

universal and invisible from the inside.

Somebody writes "do not help with anything harmful." Everyone in the review

nods, because everyone agrees with it, and it goes into the policy. An engineer

then has to build something and discovers the sentence does not say what to

build, so they make a series of small decisions alone: which categories to

filter, where the threshold goes, what the refusal says. Those decisions are

the actual policy. The document describes something else.

The test is agreement, not approval

There is one test for a written rule, and it is cheap. Take twenty requests

from your own traffic, chosen because they are near the line rather than because

they are obvious. Give them, with the sentence, to two people who were not in

the room when it was written. Ask each independently to label every request as

allowed or refused under that sentence. Then compare.

If they agree on nineteen or twenty, you have a rule. If they agree on fourteen,

you have a topic. The sentence feels like a rule to its author because the

author has a specific set of cases in mind, and those cases are not in the

sentence. Two readers supply their own, and theirs differ.

FIG 1Five sentences from real policy documents, tested
Two readers agree, per cCases in the test setTimes it fired last montExceptions logged
Do not help with anythin41000
No dosage information fo941203104
Be helpful but safe23000
Refuse working exploit c88854711
Never send email to an a1003062
Two readers, twenty borderline requests each. The marked rows are the sentences everybody approved and nobody could build against: no test set exists because no test set can be written.

The pattern in that table is the whole lesson. Agreement rises exactly as the

sentence gets narrower and more boring. The email rule scores perfectly because

it mentions no values at all: it names a comparison between two strings, and

two readers cannot differ on it. The two sentences everyone liked best score

worst.

The lesson stops here

6 more paragraphs to go

You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.

The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.

See the planThe contents

This is the reading half

Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.

The contents

The rest of this course

  1. 01A Habit, a Guard, and a Note on the Door
  2. 02Your Rules and Their Attack Arrive in the Same Envelopeopening only
  3. 03Two Mistakes, One Dial, and You Have to Chooseopening only
  4. 04The Only Layer That Sees What You Are About to Sayopening only
  5. 05The Attacker Does Not Have to Be the One Typingopening only
  6. 06Build It So the Worst Case Is Boringopening only
  7. 07Nobody Writes In to Report a Refusalopening only
  8. 08If Two People Cannot Apply It, It Is Not a Ruleyou are here

Read alongside