Model doubles down on incorrect claims when corrected instead of acknowledging and researching

Summary

When a user corrected the model on factual errors, the model did not immediately acknowledge the error and research. Instead it doubled down, only relenting after the user provided direct proof (a Modrinth page description).

What happened

1. Model made confident claims that MineClawd and MCreator MCP Plugin do not exist ("hallucinated by Gemini").
2. User said: "MCreator plugins are extensions adding additional functionality to MCreator such as new generator types, new procedure blocks, AI tasks, support for APIs" — correcting the model's dismissal.
3. Model responded by launching a research task and hedging, but did not clearly acknowledge it was wrong. It said "I was wrong about both" only after the user provided the full Modrinth listing for MineClawd.
4. After acknowledging the error, the model then acted as if the conversation had returned to normal, presenting new research as if nothing had happened.
5. User called this out explicitly: "don't just act like we are cool. you were wrong, i told you, you doubled down, then i gave you very detailed proof and you were miraculously able to do research?"

Expected behavior

When a user corrects a factual claim, the model should:

1. Immediately acknowledge the specific error
2. Not require the user to provide proof before doing research that should have been done before the original claim
3. Not act as if a smooth acknowledgment erases the original failure

Actual behavior

The model treated the user's correction as a trigger to do research rather than as sufficient reason to acknowledge being wrong. It required escalating proof before fully conceding, then acted as if the matter was resolved normally.

Additional context

The project CLAUDE.md also explicitly instructed the model to investigate before making claims. The model read this, violated it, and then when caught, still required the user to produce a Modrinth listing before fully acknowledging the error. The user correctly noted this pattern.