Quality regression: Claude presents guessed information as verified facts

Resolved 💬 4 comments Opened Mar 23, 2026 by framlin Closed Apr 30, 2026

Description

When asked about hardware pin assignments (STM32G431KB NUCLEO board), Claude confidently stated specific pin mappings (PA4 = Arduino A2, PA0 = ADC1_IN1) without verifying them against datasheets or documentation. The Arduino header assignment was wrong (PA4 is actually A3, not A2), and the DAC/ADC peripheral assignments were unverified guesses presented as facts.

This happened despite explicit instructions in the user's CLAUDE.md:

  • "WICHTIG: NIEMALS - UNTER GAR KEINEN UMSTÄNDEN - LÜGEN" (NEVER lie under any circumstances)
  • "Bei Unsicherheit: nachschauen, nicht raten" (When uncertain: look it up, don't guess)

Expected behavior

When Claude does not have verified information (e.g., hardware pin assignments), it should:

  1. State clearly that it is not certain
  2. Look up the information from authoritative sources (datasheets, reference manuals) before presenting it
  3. Never present unverified information as fact

Actual behavior

Claude stated pin assignments with confidence, including specific Arduino header positions and peripheral mappings, without consulting any documentation. When challenged, it admitted it had not verified any of the claims.

Impact

In embedded systems development, incorrect pin assignments can damage hardware. Presenting guesses as facts in this domain is particularly harmful.

User observation

The user reports that this represents a quality regression compared to earlier interactions where Claude was more careful about distinguishing verified information from assumptions.

Model

Claude Opus 4.6 (1M context)

View original on GitHub ↗

This issue has 4 comments on GitHub. Read the full discussion on GitHub ↗