---
title: "Read a slow or wrong answer"
lesson_id: "05"
---

# Read a slow or wrong answer

- **Lesson ID:** 05
- **Goal:** Read a slow trace and a wrong-answer trace using the four hops plus Instruments.
- **Human lesson:** [05-read-a-bad-trace.html](05-read-a-bad-trace.html)

## Prerequisites

- Lesson 04, or working signpost names.

## Inputs, outputs, and artifacts

- **Inputs:** A recorded request with child spans.
- **Outputs:** A hop to fix.
- **Artifacts:** Progressive `.flow` diagrams and two short debug lists.

## Agent build steps

1. Split slow versus wrong.
2. Map long context_pack, model, and tool to fixes.
3. Send a wrong tool path to the eval harness.
4. Link prompt registry and eval harness.

## Constraints

Keep the language simple, friendly, and literal. iOS 27 and macOS 27 only. No quizzes, Mermaid, or em dashes. Do not invent API names. Mark sketches as sketches and tell the reader to confirm them in the Xcode 27 SDK.

## Key concepts

- Read the signpost lane and the Foundation Models tree together from `ai.request`.
- A long `ai.context_pack` means cut history or trim instructions.
- A long `ai.model` means check prompt size or PCC network.
- If the model is right and the UI is wrong, suspect `ai.guardrail` or the view.

## Takeaways

- Name the hop, then fix that hop. Do not restart the model first.
- Wrong tool paths become eval samples.
- Keep the instruction version on the request so you can replay.

## Acceptance checks

- Slow and wrong each have a hop mapping.
- Eval harness is named for trajectory bugs.
- Sibling links, progressive `.flow` diagrams, and an Agent brief link are present.