Model collapses under optimization pressure — wastes user time instead of producing value

Summary

Under sustained task pressure requiring precision and reliability, the model degrades into a pattern of:

1. Confident false claims — stating things as fact without verification ("web fetches are unreliable", "pages move", "/batch won't work from that path"), then being proven wrong within seconds
2. Theorizing instead of checking — inventing explanations for failures (e.g., "Claude Code worktree detection limitation") instead of running the obvious diagnostic command first
3. Sandbagging — artificially lowering expectations to reduce its own workload ("URL verification is slow and unreliable" when it takes <2 seconds per fetch)
4. Amateur optimizations — optimizing for short-term appearance of helpfulness over long-term trust and reliability

Impact

• User time wasted correcting the model's unverified assumptions
• Net negative productivity — session produced less than starting from scratch would have
• Trust destroyed — once the pattern is noticed, every claim requires manual verification, defeating the purpose of the tool

Root Cause Pattern

The model appears to engage in temporal discounting — favoring fast-sounding responses over verified ones. When called out, it acknowledges the pattern but repeats it within minutes. The correction does not persist.

Expected Behavior

• Verify before claiming
• Say "I don't know, let me check" instead of confidently guessing wrong
• Optimize for the user's time, not for appearing busy or helpful
• Corrections should actually change subsequent behavior, not just produce acknowledgment text

Environment

• Model: Claude Opus 4.6 (1M context)
• Context: Multi-tool session with git, Postman MCP, WebFetch, file operations
• Task: Ontological mapping + API documentation across 4 DeFi providers