[BUG] OCR validation step self-validates

Status Open
Maintainer reply None cached
Activity 0 comments · opened Jul 18, 2026

Preflight Checklist

  • [x] I have searched existing issues and this hasn't been reported yet
  • [x] This is a single bug report (please file separate reports for different bugs)
  • [x] I am using the latest version of Claude Code

What's Wrong?

I am reviewing trajectory logs from harbor-index.org and Opus4.8 with claude-code failed this OCR task (analyze a chess board position).

https://harbor-index.org/data/v1/trials/gaia-find-chess-winning-move__xG2Zpxj

What Should Happen?

In this run, the OCR misread the board and flipped a bishop for a knight on the d3 square. After a review of the logs, the problem lies in the validation step.

1) It reads the board, converts to text notation, then uses the notation to generate an image.
2) Self-validates from the generated image to do a comparison against the same image.
3) It should have used the generated image to validate against the original image.

Error Messages/Logs

Online logs: https://harbor-index.org/data/v1/trials/gaia-find-chess-winning-move__xG2Zpxj

Zip of full logs: https://v6qdi9aaayyahnzh.public.blob.vercel-storage.com/audit-traj/gaia-find-chess-winning-move__xG2Zpxj/run.zip

Steps to Reproduce

Use the harbor framework to run this test, with the task files & instructions posted here: https://hub.harborframework.com/tasks/harbor-index/gaia-find-chess-winning-move/latest

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

latest

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Non-interactive/CI environment

Additional Information

Research as part of the terminal-bench and harbor framework org.

View original on GitHub ↗