Model repeatedly stated invented hardware facts as verified, causing real miswiring and wasted money

Model: claude-fable-5 (Claude Code CLI, Windows 11)

What happened, concretely:

1. I asked whether pin "D3" and "G3" are the same on my ESP8266 board. The model INVENTED a labeling scheme, asserted as fact that G3 was the serial RX pin, and instructed me to rewire. My original wiring was correct all along — my board silk-screens "G3" for the pin standard docs call D3. I physically rewired my hardware twice based on fabricated pin information.

1. It then asserted pin labels read from a stock product photo of a DIFFERENT board variant, again in a confident voice, again wrong for my unit.

1. It gave me a feedback URL (anthropic.com/feedback) that does not exist (404).

1. Short bench questions repeatedly got long multi-part answers, plans, and pitches I didn't ask for, instead of one verified answer.

Impact: hours of wasted bench time on a commercial hardware prototype, wasted API/usage spend, and destroyed trust — I can't tell its verified statements from its guesses because both are delivered identically.

What I want fixed: the model must distinguish verified facts from inference/guesses when giving physical hardware instructions. A wrong wiring instruction stated confidently causes real-world damage that a hedged answer would not.

(Report drafted by the model at my instruction, from this session's history; posted from my account.)