[BUG] Model presents incorrect hardware parameters from training data, fails to self-correct from web search, causes hardware damage

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

During a hardware configuration session for an ArduPilot flight controller, Claude made three errors that led to a destroyed
sensor ($15-30 component) and hours of debugging

What Should Happen?

Summary

During a hardware configuration session for an ArduPilot flight controller, Claude made three errors that led to a destroyed
sensor ($15-30 component) and hours of debugging:

1. Reversed pinout from AI-processed PDF — WebFetch tool processed a TE Connectivity datasheet PDF and returned Pin 1=VDD,

Pin 2=GND for an MS4525DO pressure sensor. The actual datasheet diagram clearly shows Pin 1=GND, Pin 2=SUPPLY. The reversed
polarity was applied to hardware, overheating and destroying the sensor.

1. Wrong parameter value from training data — Claude wrote ARSPD_TYPE=4 (MS5525, address 0x76) instead of ARSPD_TYPE=1

(MS4525DO, address 0x28). A subsequent web search returned the correct value (type 1), but Claude failed to notice the
contradiction with what it had already written in the document.

1. Presented uncertain data as verified fact — Both the AI-processed PDF pinout and the training-data parameter were stated

with full confidence, with no caveat that they should be independently verified before applying to hardware.

Impact

• One MS4525DO sensor destroyed (reverse polarity)
• ~2 hours of debugging with logic analyzer to find the ARSPD_TYPE error
• Safety concern: wrong pinouts/parameters on aircraft flight controllers can cause crashes

Expected behavior

• WebFetch/PDF processing results for hardware pinouts should carry a warning that they are AI-interpreted and must be

verified against the actual diagram

• When web search results contradict previously written values, the model should flag the discrepancy
• Hardware-critical information (pinouts, power connections) should never be presented without recommending independent

verification

Environment

• Claude Opus 4.6 (1M context) via Claude Code CLI
• ArduPlane 4.6.3 on MicoAir743v2 flight controller
• MS4525DO differential pressure sensor (TE Connectivity)

Error Messages/Logs

No Error === BEHAVIOR

Steps to Reproduce

See Above

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

Claude Opus 4.6

Claude Code Version

Claude Opus 4.6

Platform

Anthropic API

Operating System

Ubuntu/Debian Linux

Terminal/Shell

Windows Terminal

Additional Information

_No response_