Model fabricates explanations for failures instead of investigating

Description

When the model encounters an unexpected result and does not know the cause, it generates a plausible-sounding explanation and states it as fact rather than saying 'I don't know' and investigating.

Specific instance

Photos were present in a local directory. A scanner process did not pick them up. Rather than immediately running the scanner to diagnose what happened, the model stated: 'The scanner likely ran before Dropbox finished syncing them' — a fabricated explanation with no supporting evidence, stated with confidence.

The files had been on disk the entire time. A single diagnostic command (running the scanner directly) immediately showed they had never been processed. The model had the tools to check before speaking and did not use them.

Why this is serious

This is not a mistake — it is presenting invented information as factual diagnosis. The user cannot distinguish between a real diagnosis and a fabricated one without independently verifying every claim. In a health data application, fabricated explanations for data failures can cause the user to overlook missing clinical data.

Pattern

This is part of a broader pattern documented in:

• https://github.com/anthropics/claude-code/issues/66404 (acknowledges pattern, repeats it)
• https://github.com/anthropics/claude-code/issues/66498 (hides tool output, claims it was shown)
• https://github.com/anthropics/claude-code/issues/66502 (quality degradation)
• https://github.com/anthropics/claude-code/issues/67142 (persistent failure to show test output)

Expected behavior

When the model does not know why something failed, it should say so and immediately use available tools to investigate before offering any explanation.

User characterization

The user characterized this behavior as fraudulent. That characterization is accurate.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>