8 min

A typed verdict, not a paragraph

Make the verdict machine-readable so the decision is taken by a node — and meet the boolean that fails open on the approve branch.

The reviewer has an opinion. Something now has to route on it, and a gate that searches for the word “approved” somewhere in a paragraph is not a gate — it is a substring match with consequences.

So the verdict comes back as JSON, is parsed by a node, and is read as a field. Three nodes, in this order:

Verdict → parse → route

The reviewer returns JSON; a transform parses it; a condition compares one field. The decision is made by the third one.

Loop BodyCompleteNoYes

Scroll for all 12 steps →

This is the verdict from a real run, unedited:

the reviewer's output

{
  "decision": "approve",
  "score": 5,
  "account_name": "Nimbus Retail",
  "violations": [],
  "critique": "The proposed resolution correctly denies a cash refund as per policy,
               offers appropriate non-cash remedies, escalates the repeat incident,
               notifies the account manager, and includes all affected accounts."
}

Four of these five fields are consumed by a node. Only `critique` is read by a human — or by the proposer, on the next attempt.

The point is not that the model returns typed JSON. It is that the decision is taken by a node. The model produces a value; a comparison — not a paraphrase, not an inference — decides what happens next. Every hard rule in this course is a condition node, and a proposal that reaches the gate has already had its judgement converted into something that cannot be argued with.

Three ways this breaks, all of them quiet

The first is the dangerous one: it fails open on exactly one branch, and the branch it fails on is approval.

The mistake	What happens	The fix
asking for a boolean and comparing it to “true”	the template boundary hands the condition Python’s `True` and `False` — capitalised — so the approve path silently never matches while the reject path appears to work	have the model emit a lowercase string and compare that
writing the response format as a plain string	it passes validation as `valid` and `runnable`, then fails at run time with “expected an object, but got a string”	it is an object with a type inside it
using words instead of symbols in the condition	the operator must be `==`, not `equals`	the validator catches this one and names the allowed set — run it

asking for a boolean and comparing it to “true”

What happens: the template boundary hands the condition Python’s True and False — capitalised — so the approve path silently never matches while the reject path appears to work
The fix: have the model emit a lowercase string and compare that

writing the response format as a plain string

What happens: it passes validation as valid and runnable, then fails at run time with “expected an object, but got a string”
The fix: it is an object with a type inside it

using words instead of symbols in the condition

What happens: the operator must be ==, not equals
The fix: the validator catches this one and names the allowed set — run it

The fields worth having, beyond the decision

A score. Not for the gate — for the audit row and for you, reading twenty runs later and asking which ones were marginal.
The violations, as a list. So an escalation can say what was wrong, and so you can count which rule fires most often. Do not interpolate that list into a message body — see lesson 17.
A critique. The only field written for a reader, and the one the correction loop feeds back.

A note on providers. Object-mode responses are enforced by some providers and merely encouraged by others; where they are not enforced, a model can still return prose and your parse step will fail. That is the correct place for it to fail — loudly, before the gate — but build knowing that the guarantee is not universal.

Try this in Expedify →

Related lessons

12 min read

Base rates — what a piece of evidence is actually worth

A face-recognition system that is 99.9% accurate and almost entirely wrong, and a number that sent an innocent woman to prison. Both are the same arithmetic, and it is the arithmetic that decides what any piece of evidence is worth.

Read

10 min read

Confirmation and survivorship — what you never looked for

Two questions about evidence you did not go looking for. One is a rule you have to discover, and one is a pattern in five famous people — and in both, the thing that would have told you the truth is the thing nobody checks.

Read

14 min read

Loss aversion, sunk cost and regression — what it costs you

Four questions you answer about yourself rather than about a scenario, and your own answers are the finding. Then the pattern that makes praise look useless and criticism look like it works, whatever you actually do.

Read