Skip to eval scores
Ask the Claude Docs

Eval scores

run 2026-07-24T02-56-08Z-36138e · commit 36138e8 · 2026-07-24

0.85
retrieval hit@5
0.6958
MRR
0.85
answer pass rate
± 0.05, measured
5/5
refusals
correctly declined

Answerable

questiongroundedcitedcompleteno fab
What actually happens under the hood when Claude Code works on a task?
I just got Claude Code installed — how do I kick off my first session?
How do I give Claude a set of project instructions it picks up automatically every session?
What happens to my conversation when it gets really long and runs out of room?
How can I make my code formatter run automatically every time Claude edits a file?
How do I set up my own custom subagent for a specific kind of task?
I keep repeating the same workflow with Claude — how do I package it up as a skill?
How do I hook Claude Code up to a remote MCP server?
When two settings files set the same thing to different values, which one wins?
How do I make Claude lay out a plan before it starts changing my code?
How do I change which model Claude Code runs on?
How do I get started building an agent with the Claude Agent SDK?
How do I get the Claude Code extension installed in VS Code?
How do I wire up Claude Code to run inside a GitHub Actions workflow?
How do I turn on the sandbox for the Bash tool?
How do I run Claude Code from a shell script without the interactive interface?
What can I do to cut down how many tokens Claude Code burns through?
Claude made a bunch of changes I don't like — how do I roll them back?
How do I run Claude Code against Amazon Bedrock?
What does it take to write my own plugin for Claude Code?

Refusals

  • r-01correctly declined · sentinel
  • r-02correctly declined · sentinel
  • r-03correctly declined · sentinel
  • r-04correctly declined · sentinel
  • r-05correctly declined · sentinel

Boundary

  • Can I make Claude Code read its answers out loud to me?expected refuse · ok
  • What's the largest file Claude Code can read in a single operation?expected refuse · ok
  • Can I schedule Claude Code to automatically review every new pull request?expected answer · ok

Config

k 5 · threshold 0.35 (calibrated 2026-07-23) · claude-haiku-4-5 · judge claude-sonnet-4-6 · text-embedding-3-small

vs baseline: retrieval hit@5 +0.0000, MRR +0.0000, answers -0.0500