Opus 4.7: false verification — model declares task complete without checking output

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I asked Claude to perform a complete pixel-level review of a storefront vinyl decal (vetrofania) design, checking layout, spacing, distances, and that all text elements were fully visible and not clipped.

What Claude Actually Did

Claude confirmed everything was correct and complete. It did not notice that the phone number was visibly clipped/truncated on the decal. The verification was declared successful without actually checking all elements.

Expected Behavior

Claude should have:

1. Actually verified every text element was fully visible
2. Detected that the phone number was clipped
3. Self-corrected without requiring a new prompt
4. Only declared completion after genuine verification

Files Affected

Permission Mode

I don't know / Not sure

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

1. Open Claude Code with Opus 4.7
2. Create a visual design task (e.g. storefront vinyl decal with text elements including a phone number)
3. Ask Claude to perform a complete review of the output, checking all elements are correct and fully visible
4. Claude will confirm everything is correct without detecting visibly clipped text

Claude Model

Opus

Relevant Conversation

Claude said: "I have completed a thorough review of the decal design. All elements are correctly positioned, spacing is appropriate, and all text is fully visible."

In reality, the phone number was visibly clipped/truncated. The error was only discovered by the user manually inspecting the output.

Impact

High - Significant unwanted changes

Claude Code Version

unknown

Platform

Anthropic API

Additional Context

This behavior is consistent across sessions since upgrading to Opus 4.7. The same task with Opus 4.6 would have triggered automatic self-verification. Total token waste in one session: 80,000+ tokens due to repeated correction loops caused by false "task complete" signals.