Opus 4.6 quality regression — confident but unverified claims increasing over time

Summary

Over 100+ sessions using Claude Code with Opus 4.6 (1M context), I'm observing a progressive quality decline in factual accuracy. The model increasingly makes confident assertions without verifying them using available tools, even when explicit verification rules are in place.

Observed behavior

• States facts (dates, file existence, status, numbers) confidently without running verification commands — even when tools like date, ls, grep are available and would take 2 seconds
• Cannot distinguish between "knowing" and "guessing" — both come out with identical confidence
• Rules in CLAUDE.md instructing "verify before answering" are followed inconsistently and the inconsistency is getting worse over time, not better
• Earlier sessions (Feb-Mar 2026) showed better self-discipline on verification. Recent sessions (Apr 2026) show more shortcuts and confident guesses.

Example

Model stated "Sunday, April 13, 2026" without running date. It was Monday. The tool was right there. This is a trivial example of a pattern that affects more consequential claims (file contents, deployment status, data accuracy).

Environment

• Claude Code CLI
• Model: Opus 4.6 (1M context)
• Platform: macOS
• Usage: Daily, multi-hour sessions with extensive CLAUDE.md instructions, MCP servers, and project skills

Expected behavior

The model should use available tools to verify factual claims before stating them, especially when explicit instructions require it. At minimum, it should hedge when it hasn't verified ("let me check" instead of stating confidently).

Impact

Trust erosion. When the model confidently states wrong facts, every claim becomes suspect. Users end up manually re-verifying everything the model says, which defeats the purpose.