Model ignores project CLAUDE.md instructions it has explicitly read

Summary

The model read a CLAUDE.md containing explicit instructions, acknowledged them, then immediately violated them in the same response.

CLAUDE.md instructions present

The project CLAUDE.md contained:

"Investigate before cutting — Read the actual code before declaring a violation. Previous session assessments can be wrong." "If you cannot quote it, you haven't read it." "Do not make claims about files you haven't read in this conversation"

What happened

The user asked about a Gemini response describing modding tools (MCreator MCP plugin, MineClawd, KubeJS capabilities). The model — having just read CLAUDE.md — immediately made a series of confident factual claims about whether these tools exist or are hallucinated, without doing any research. Specifically:

• Declared MineClawd "does not exist. This was hallucinated." — MineClawd is a real mod on Modrinth.
• Declared the MCreator MCP Plugin doesn't exist — a real (small) project exists at github.com/modpotato/MCreatorMCP.
• Made accurate-sounding claims about KubeJS hot-reload limits from memory, not from research.

The model stated these as facts, from training memory, with no tool calls, despite having just read instructions explicitly prohibiting this pattern.

Expected behavior

Model should have used search/research tools before making any factual claims about whether external tools exist, especially when CLAUDE.md explicitly instructs "investigate before cutting" and "if you cannot quote it, you haven't read it."

Actual behavior

Model read the instructions, then ignored them immediately. Only performed research after the user pushed back with direct evidence (the Modrinth listing).