[Bug] Claude generates fabricated content instead of requesting clarification

Bug Description

Claude Code repeatedly presents unverified assumptions as facts, acts on them, and only checks authoritative sources after being challenged.

Failure pattern:

• invents facts instead of verifying them;
• treats local/project context as ground truth when external verification is needed;
• makes confident claims about current external systems without checking official sources;
• modifies files based on guessed information.

When challenged, Claude ran the searches it should have run upfront and found multiple fabrications in its own output. A follow-up audit surfaced additional unsupported claims. The verification tools work when invoked, but they are not invoked proactively.

This occurred with Opus 4.7, effort: max, so max effort did not prevent avoidable factual fabrication. The issue appears to be verification-trigger behavior, not simply reasoning depth.

Expected Behavior

Claude Code should verify external factual claims before using them, clearly separate verified facts from assumptions/unknowns, and avoid modifying files based on unverified information without warning.

Actual Behavior

Claude Code made confident false claims, wrote them into files, and corrected itself only after direct challenge.

Impact

Syntactically valid changes may contain invented or unsupported facts, appearing authoritative enough for users to ship wrong data.
Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.138
• Feedback ID: 96a1a4a8-6cfd-4179-9cd4-a972421360a0

Errors

[]