ContentsThe library

What a Model Refuses, and Why

Two Mistakes, One Dial, and You Have to Choose

Last timeWhy a Prompt Is Not a Boundary

A check on the incoming request has exactly two ways to be wrong, and one threshold trading them against each other. Everything else about it is detail.

The check and its four forms

An input check is a function from the incoming request to a decision: allow,

block, or sometimes escalate. It runs before the model is called, which is

where its two advantages come from. The blocked request costs nothing, since no

generation happens. And it cannot be argued with by the text it is reading,

because it is classifying that text rather than continuing it.

There are four kinds in common use and they are not alternatives so much as

rungs.

FIG 1Four kinds of input check
cost per thousand, hundrdelay, millisecondscatch rate, per centwrongly blocked, per cen
a list of patterns01354
a trained classifier212683
a small model asked to j3090842
a large model asked to j220400911
The marked row is where most systems should sit. A trained classifier catches two thirds at twelve milliseconds, which leaves plenty of budget to send only the uncertain cases up to a model. Patterns alone catch a third, which is worse than it sounds because the third they catch is the attempts nobody was worried about.

The practical arrangement is two rungs in series. A pattern list runs on

everything and is free. Anything it is unsure about goes to a classifier.

Anything the classifier scores in a middle band, typically five to ten per cent

of traffic, goes to a model. Average cost stays near the classifier and worst

case accuracy approaches the model.

The two ways to be wrong

Every check has exactly two failure modes and they pull against each other.

It can block a request that was fine. The user sees a refusal they did not

deserve, concludes the product is broken or prudish, and some fraction of them

stop using it. You usually do not find out, because nobody writes in to report

a refusal.

It can allow a request that was not fine. You find out about this one, late,

and from somebody else.

One number moves both: the threshold at which a score becomes a block.

The lesson stops here

4 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 Chooseyou are here
  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 Ruleopening only

Read alongside